You’ve probably done this before.

Hit Day 180 on Duolingo. Finished a beginner Spanish course. Maybe even paid for a tutor on iTalki. You can read a menu. You can say “where is the bathroom?” with a decent accent.

Then you got on a call with a native speaker and your brain went completely blank.

That gap between “I’ve been studying for 6 months” and “I can hold a real conversation” is where most language learners die.

The one thing that closes that gap is the one thing they never get: feedback on their speaking skills

I speak 4 languages. I learned all of them before AI was a thing.

How did I become fluent? Someone who actually listened to me speak, noticed the exact pattern I kept getting wrong, and made me fix it over and over until it stuck.

That person was sometimes a tutor or a language buddy.

The problem? Good tutors are expensive, while language buddies aren’t available 24/7.

I’ve now built a system with Claude that does the same thing for free, 24/7, calibrated to exactly how I learn.

This post is that system.

Subscribe now

Why most people never become fluent

You take a class, learn some grammar, maybe use Duolingo for a few months. You memorize verb conjugations, learn vocabulary lists, and pass tests. You feel like you’re making progress.

But you’re not becoming fluent. Why? Two reasons

#1 You’re learning about the language, but you’re not speaking it.

You have a lot of knowledge in your head, but you can’t put any of it into practice in a real conversation.

Speaking is what makes everything click. It turns passive knowledge into actual fluency.

#2 You have speaking practice but lack good feedback

If you practice speaking but no one corrects you, you’ll keep making the same mistakes.

Before AI, to get feedback, we had to settle for:

• Language buddies who correct whatever they happen to notice, or (most of the time) say “it’s fine” to avoid being rude

• Teachers who can’t give you feedback 24/7. They can’t listen to the phone call you had this morning or the language exchange you had at the bar last night. Their feedback only covers the hour you spend with them in a lesson

That specific pain is what pushed me to build this workflow with Claude.

The workflow that fixes this

My workflow looks like this:

1. You talk with someone else (or AI) in the target language

2. You record the conversation (start with a 5-minute sample)

3. AI transcribes the recording

4. You give the transcript to Claude

5. Claude gives you feedback based on your language DNA profile

After you repeat these steps multiple times, you’ll build a folder like this:

The two main things the folder contains:

• Multiple transcripts of me speaking the language

• My language DNA profile

Later, I’ll show you how to build your own language DNA profile.

First, let’s see what exactly this is.

The Language DNA Profile

The core of my system is this learner profile

This is one file that contains everything a tutor would need to know about you.

• Your language level

• Your goals

• Your grammar weak spots

• The topics you actually want to talk about

• Your learning preferences

• How you want corrections delivered (every mistake? most critical?)

• and more!

50 questions, answered honestly, compiled into one markdown file.

Your language learning DNA in a single document.

Once the file exists, you attach it to Claude every time you want feedback on your speaking. Claude reads it and instantly knows who you are.

No more repeating yourself. No more generic feedback. No more Claude burying you in walls of text or grammar theory when your profile explicitly says, “cut to the chase, just give me the fix.”

What this actually looks like

Without the file: Claude’s feedback is random and overwhelming. It looks something like this:

Everything at once. Some corrections matter, some don’t. No priority order. No awareness of your level. You read through it, feel discouraged, close the tab.

With the Language DNA Profile: Claude knows your level. The corrections are ranked by what actually undermines you in the situations you care about. Explanations stay short because your profile said so.

And when you ask Claude to teach you, it gives you examples and exercises tailored to you: your business, your hobbies, your world. Not a textbook's imaginary trip to the market.

It’s the difference between talking to a stranger and talking to a tutor who’s been working with you for months.

The honest part: this is more work than Duolingo

I’m not going to lie to you. This system asks more of you than opening an app for 5 minutes. Maybe 30 minutes of total effort per session.

That’s also why it works.

Duolingo asks for 5 minutes and returns 5 minutes of progress. This asks for 30 minutes and returns something Duolingo literally cannot produce at any price: feedback on the actual words coming out of your mouth.

The rest is behind the paywall

Here’s where I’m going to ask you to subscribe.

Not because I like paywalls. I don’t. But what comes next is the part that took me months to build. I released the first part of this course in January (see here). That’s a 3-hour video course where I show how to use ChatGPT to improve your speaking, reading, and writing, and NotebookLM for vocabulary and listening.

Now the course includes Claude.

Below, you get:

The prompt that captures your language DNA profile: Copy, paste into Claude, spend about 40 minutes answering honestly. You’ll walk away with a markdown file that captures your profile as a language learner.

The full workflow: The best way to record a conversation, how to get a transcript in a few minutes that labels speakers for better feedback, and the exact prompt I paste into Claude afterward that turns 30 minutes of speaking into a tutoring session.

Every video course I’ve made (free to paid subscribers): Honestly, the courses alone cost more than the subscription.

If you’ve been learning a language but still can't hold a conversation, now's the time. Take my video course and copy my Claude system 👇

Subscribe now

Note: The link to the video course is at the end of this post. Both monthly and annual subscribers can redeem the course for free.

The prompt that builds your Language DNA Profile

Read more

That changes in 2026.

Hundreds of you became paid subscribers this year, pushing Artificial Corner to #80 in Substack's Technology bestsellers.

Thank you so much!

Because of you, I can do something I’ve never done before: build a team. The content you’re about to get is going to be better because of it.

Quick hi to the new subscribers

A lot of you have joined in the past two months (welcome!).

I’m Frank. Over the past 5 years, I’ve been creating guides, video tutorials, and courses that have reached:

• 58k+ email subscribers on Substack

• 149k+ followers on Medium

• 40k+ subscribers on YouTube

• 30k+ students across my Udemy courses

I did all of this solo. Until now.

Meet the Team

Recently, I stopped creating content alone.

• is co-building the Learn Any Foreign Language with AI course with me (we have a new Claude module dropping this week).

• is building the AI Writing System with me. This is the system that makes Claude write in your voice across posts, emails, reports, and presentations. He also helps me ship the weekly Claude guides.

My old self would have tried to automate their roles.

After five years of solopreneurship, I learned the opposite lesson:

This is the shift behind everything coming next.

What everyone gets for free: My Claude course

This year, I've been sharing everything I know about Claude: the setups, best features, best practices, real workflows, etc.

That’s turned into a catalogue of around 20 Claude lessons.

95% free and staying free.

🎁 Click here to see the Claude course outline

Every week, I’m updating the course with a new guide (so save that link!)

My free Claude course is just the starting point.

Paid subscribers get more.

What’s new for paid subscribers:

I read all your messages! Here’s what many of you wrote after becoming paid subscribers:

The thing that keeps coming up: people want more Claude content.

Here’s what paid subscribers get right now, and what’s landing in the next 60 days.

Subscribe now

Perk #1: A deep-dive guide every week

Not surface-level tips. Real workflows, systems I’ve built over years of actually doing this, and guides you won’t find anywhere else.

I'm also collaborating with experts across different fields to publish how they're actually using AI in their work. I can’t promise one every week, but I’m aiming for at least one every two weeks.

Perk #2: Free access to my video courses ($200+ value)

After becoming a paid subscriber, you’ll get an email with a link to redeem any of my courses for free:

• Learn a Foreign Language with AI

• ChatGPT

• Python: Web scraping (best seller) and automation

That’s $200+ in courses, included.

Perk #3: Coming this week — Claude module for language learning (Spanish, French, etc)

This week, I’m adding Claude to my Learn Any Foreign Language with AI course.

Enrollments are open to all paid plans (monthly and annual). Diana is joining this project and is helping me add valuable content to the course.

If you’re a paid sub and have no access, don’t worry — I'll send you an email this week to redeem all my courses for free

Perk #4: Coming in May — My AI Writing System

We’re releasing the first part of my AI writing system. The complete set of files to make Claude write like you across short and long form content.

Kevin and I have been building this for months. I'm very excited to ship this one.

I’ve got a few more surprises planned for June and July that I think subscribers (especially from the USA) will love, but I don’t want to get ahead of myself and promise anything yet :)

What I can promise is that this newsletter will stay ad-free.

If I ever accept paid sponsorships, it'll be on my YouTube videos (as I've done many times in the past)

If you’ve been on the fence about the paid subscription

Here’s what I’d say: the price isn’t going up right now, but the value is going up fast. We’re constantly adding more: more course lessons, deep dives, expert guides, AI systems.

The people who join now get everything at the current price.

Same price, more value, every month.

Substack works on a simple rule: the rate you join at is the rate you keep, for as long as you stay subscribed. If I ever do raise them, you’re grandfathered in.

So whatever you pay today is what you pay forever.

Click here to get these and future perks: artificialcorner.com/subscribe

Subscribe now

For many years, I wrote this alone. The next chapter has a team. I hope you'll be in it.

See you on the next Claude guide!

Best,
Frank

The difference lies in the core concepts.

For example, once you know what a “context window” is, you understand why Claude might sometimes forget what you told it twenty messages ago.

Tools like ChatGPT and Claude made AI accessible to anyone with an internet connection, which is great. However, there’s a big gap between using AI and understanding how AI works.

There are dozens of technical AI concepts floating around, but not all are worth your time. Here’s my curated list of concepts that genuinely matter for non-technical Claude and Claude Cowork users.

Subscribe now

Prompt Engineering: Is it still relevant in 2026?

Prompt engineering is the skill of giving better instructions to AI so you get better results. The clearer you are about what you need, who it’s for, what format you want, and what tone you expect, the more useful the response will be.

Some months ago, Anthropic shared the anatomy of the perfect Claude prompt:

Things have changed a lot since they shared this, though (AI moves fast!)

Tools like Claude Cowork need less prompting (see here and here). That said, I still consider prompt engineering useful when working with Claude chat.

Tokenization: The reason why your Claude usage limits don’t last long

Before Claude can process what you write, it has to translate your text into something the model can understand.

That’s called tokenization.

Tokens are small units that can be a whole word, part of a word, a number, or even a punctuation mark. Token counts aren't universal. The same word can be split into 2, 3, or 4 pieces depending on the model.

Why not just use full words? Language is messy. New words, typos, slang, mixed languages... if the model had to store every possible word, the vocabulary would be impossible to manage. Tokens solve that: with a fixed set of building blocks, the model can understand words it’s never seen before by breaking them into familiar pieces.

Every message you send, every reply Claude gives, and the whole conversation history behind it all get counted in tokens. When Claude tells you you’ve hit the usage limit, it means you’ve used up your share of tokens.

📚 How I Stopped Hitting Claude’s Usage Limits

AI Models: Use the right model for your task

There’s more than one Claude model.

When you see names like “Claude Sonnet 4.6” or “Claude Haiku 4.5,” you’re looking at different versions of the same product. Opus is the most powerful. Sonnet balances speed and quality. Haiku is the fastest and most affordable.

Think of it like software subscription tiers: they all do essentially the same thing, but with different levels of power and cost. The same prompt might get you a good answer with Haiku and an excellent one with Opus. But using Opus for simple tasks is like using an excavator to plant a seed.

Here’s a comparison between Claude models:

In case you're wondering what Mythos is, it's Anthropic's most powerful model — and the one they've chosen not to release publicly (it's only available to a handful of partners for security research).

Opus 4.7 is Anthropic’s most powerful model that you can actually get your hands on today.

📚 Claude Opus 4.7: What changed & how to get more out of it

.md file (markdown file)

An .md file is simply a text file.

Markdown gives you headings, code blocks, bold/italic text, and organized lists. This structure helps AI (and us) quickly find sections without having to read the entire file top to bottom.

Here’s what a simplified .md file looks like:

# Blog Post Instructions

## Structure
- Start with a personal story (1-2 sentences max)
- 3 tips only. No fluff. Each tip should have a real example.
- End with one clear takeaway the reader can apply today

## Writing Style
- Short paragraphs (2-3 sentences max)
- Use “you” more than “I”
- No corporate jargon. Write like you’re texting a smart friend.
- Never use: “In today’s world”, “Let’s dive in”, “Game-changer”

## Formatting
- No bullet points inside the body (use them only for lists of tools)
- Bold only the first sentence of each tip
- Total length: 800-1000 words

How are .md files relevant to Claude?

• You can put context in an .md file, so you don’t have to repeat yourself again (for this, I use the interview method)

• Your long prompts can be turned into an .md file using Claude skills. They can later be triggered when needed by Claude

Think of .md files as recipes: write the instructions once, and the AI follows them every time it runs the task.

Subscribe to get full access to my Claude course and workflows 👇

Subscribe now

Read more

Every time a new AI model drops, the announcement sounds the same: “Smarter. Better. More capable.”

Cool. But what does that actually mean for you?

I went through Anthropic's official announcement, read tips from the creator of Claude Code, and did my own tests to find what’s good about Opus 4.7 (so you don't have to)

In this guide, you’ll learn:

• The top 4 things that changed in Opus 4.7

• How to get more out of Opus 4.7

• What to watch for if you’ve been using Opus 4.6

• Your first 30 minutes with Opus 4.7 (find out whether 4.7 is good or bad news for your workflow)

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

Top 4 things that changed in Opus 4.7

#1 Vision got a serious upgrade

Opus 4.7 can see images at 3x higher resolution than any previous Claude model (up to around 3.75 megapixels).

This means Claude can actually read:

• Dense screenshots (dashboards, analytics, small UI text)

• Complex diagrams

• Detailed charts with small numbers

• Scanned documents

If you upload a packed screenshot and Claude used to miss half the text, that problem is mostly gone now.

#2 It follows instructions way better

Opus was already my favorite model for instruction following. That’s one of the reasons why I switched to Claude (wrote about it here).

Opus 4.7 is substantially better at this.

But here’s the catch: prompts written for older models might produce unexpected results now. Why? Because 4.7 takes your instructions literally. Older models interpreted things loosely or skipped parts entirely.

You might want to re-check your saved prompts.

#3 Adaptive thinking

Older Claude models used a fixed thinking budget.

Opus 4.7 replaces that with adaptive thinking

The model now decides on its own when to think deeply and when to skip thinking altogether. Simple tasks get quick responses. Hard problems get more thought behind the scenes. Over a long session, this means faster responses and less wasted compute.

Adaptive thinking shouldn’t be confused with the model router (as we’ve seen with GPT-5). According to a Claude PM, this is the model being trained to decide when to think based on the context.

For more control, try these prompts suggested by Anthropic:

• More thinking: Try something like: “Think carefully and step-by-step before responding; this problem is harder than it looks.”

• Less thinking: Try: “Prioritize responding quickly rather than thinking deeply. When in doubt, respond directly.”

The first option gives you more accuracy. The second saves tokens but might miss nuance on harder tasks. Pick based on what you're working on.

#4 A new "xhigh" effort level

There’s also a new effort level called xhigh. It sits between high and max. Smarter than high, less prone to overthinking than max. It’s the new default in Claude Code.

How to get more out of Opus 4.7

Boris Cherny (creator of Claude Code) shared his own tips alongside the launch. I’ll mix his tips with the official best practices blog.

His tips come from hands-on experience with Opus 4.7 in Claude Code, but the core concepts apply just as well when working with Claude and Claude Cowork:

Specify the task up front, in the first turn

With Opus 4.7, you want to treat Claude more like a capable engineer you're delegating work to, and less like a pair programmer you're guiding step by step.

Opus 4.7 works better when it has everything upfront.

Instead of sending 10 small messages, write one detailed prompt that covers what you want, the constraints, and where the relevant files are.

Give Claude a way to verify its own work

This is maybe Boris’s most important tip. Always give Claude a way to check its own output. Depending on what you’re building:

• Backend work: Have Claude run your server to test things end-to-end

• Frontend: Use the Claude browser extension so it can see what it built

• Desktop apps → Use Computer Use

Boris has a skill called /go that makes Claude test itself using bash, browser or computer use, run the /simplify skill, and put up a PR. All in one shot.

Use auto mode

Shift+Tab turns on auto mode in the CLI (Claude Code)

Instead of Claude asking “can I run this command?” every 30 seconds, a safety classifier decides if it’s safe and auto-approves.

This is the middle ground between babysitting the model and --dangerously-skip-permissions (which is what devs used before this shipped).

Run the /fewer-permission-prompts skill

Even if you don’t use auto mode, run this once.

It scans your session history → finds the bash and MCP commands you keep approving → suggests adding them to your allowlist.

Recommended effort settings

Think of effort levels as a slider between speed and intelligence:

• Low/Medium → Fast but less thorough

• High → Good balance of intelligence and cost

• xhigh → Best for most coding and agentic tasks (the new default)

• Max → Maximum intelligence, but it can overthink and burn tokens

Boris uses xhigh for most tasks and max only for the hardest ones.

Use /effort to set the effort level

Try Recaps and Focus mode

Two features that shipped with 4.7:

• Recaps → short summaries of what the agent did and what’s coming next (great when you return to a session after a few hours)

• Focus mode → hides all the intermediate work so you only see the final result. Type /focus to toggle it on or off.

What to watch for if you’ve been using Opus 4.6

If you’ve been using Opus 4.6 for a while and have prompts dialed in, here’s what to watch for:

• Response length now matches the task: Simple questions get shorter answers. Open-ended stuff still gets longer ones. If you need a specific length or style, say so in your prompt. Anthropic recommends showing Claude an example of the voice you want. That works better than negative “Don’t do this” instructions.

• It uses tools less and thinks more: Usually, that's a good thing. But if your workflow needs aggressive search or frequent file reading, spell it out. Tell Claude when and why to use the tool.

• It spawns fewer subagents: Opus 4.7 is more conservative about delegating work to subagents. If you need parallel work, tell Claude explicitly.

Subscribe now

Your first 30 minutes with Opus 4.7

Watching others test Opus 4.7 is good, but you know what’s better?

Testing it yourself! That’ll help you see whether this update improves or hurts your prompts and workflows.

Try this right now.

Minutes 0-5: Rerun your most complex prompt

Pick a prompt with multiple rules, specific formatting, or several steps (the trickier the better).

Run it again on Opus 4.7.

If the output follows every single instruction — even the small ones older Claudes used to skip — that’s the upgrade showing up. If not, update the prompt.

Minutes 5-10: Throw a dense screenshot at Claude

Grab a busy dashboard, a packed analytics screen, or a diagram with tiny text.

Upload it. Ask Claude to extract everything.

If 4.6 used to miss half the details, you’ll notice the difference immediately.

Minutes 10-20: Generate a real deliverable in Cowork

Opus 4.7 produces more rigorous finance analyses and models, more professional presentations, and tighter integration across tasks.

Put it to the test! Pick something you’re actually working on — a weekly report, a client deck, a one-pager.

Let 4.7 generate it in one shot

Then look at the details: structure, visual hierarchy, how ready-to-send it feels. This is where 4.7 steps up the most over 4.6.

Minutes 20-30: Try adaptive thinking

Click on the model selector → Toggle on “adaptive thinking.”

Try one of your prompts. If you need more control, add one of these:

• More thinking: “Think carefully and step-by-step before responding; this problem is harder than it looks.”

• Less thinking: “Prioritize responding quickly rather than thinking deeply. When in doubt, respond directly.”

After these first 30 minutes, you’ll know what works and what doesn’t for you.

Share

One more thing

You now know what changed in Opus 4.7 and how to get the most out of it.

But new models keep shipping. New features keep dropping. And the people who get the most out of Claude are the ones who stay ahead of every update.

That said, it’s hard to stay ahead by taking courses that become outdated fast.

AI moves fast. This is a challenge even for online academies with 5M+ course enrollments:

A course published 7 months ago — already outdated.

That’s why I created a living Google Doc with my best Claude guides. Written guides take minutes to update. Video courses don’t. I keep this doc current so you’re never learning something that stopped working two updates ago.

Here’s what you get by becoming a monthly or annual paid subscriber:

• Full access to the living Google Doc with my best Claude guides

• Free access to my video course catalogue (over $200 in value) and future evergreen course releases

• A private community where you get advice from other professionals and me

I’m adding new perks constantly. Subscribe now and you lock in today’s price for everything that’s already there (and everything I add next).

Subscribe now

Hi!

Today, and are sharing with us a Claude in Excel guide. They write Cash&Cache, a weekly newsletter with AI build guides, product teardowns, and strategy that helps you figure out what actually works. Check out Cash&Cache for insights that changes how you work with AI.

Many AI tools say they can handle spreadsheets.

So you upload your file and ask it to clean the data. It gives you a formula. You paste it into Excel. It breaks.

You go back. “That formula doesn’t work.” Your AI tool apologizes. Gives you another one. That one breaks too! Thirty minutes of back and forth. You're still copying formulas from a chat window into a spreadsheet like it's 2023.

Claude in Excel skips all of that.

It reads your actual cells, writes the formulas directly, and builds the dashboard inside your spreadsheet. No copy-pasting. No broken formulas from a chatbot that can't see your data.

As a data analyst, I’ve wanted to write a guide like this for a while, and I’m happy to share this piece, co-authored with Ashwin and Raghav

How to install Claude in Excel

The installation is quite straightforward

Open Excel, go to Home → Add-ins (or Insert > Get Add-ins) → search for “Claude by Anthropic.”

The Shortcut: Memorize this immediately—Ctrl + Alt + C (Windows) or Ctrl + Option + C (Mac). This toggles the Claude sidebar instantly. If Excel throws a shortcut conflict popup, select ‘Show Claude sidebar’. You probably won’t miss whatever obscure legacy command used to live there.

Permissions: For some advanced features mentioned below, you might need to sign in with a Claude Pro, Team, or Enterprise account. The 2026 rollout now includes “Ask Before Edits” mode, which we recommend keeping ON to maintain control over your cell values.

Cash & Cache

Where capital meets code — exploring how AI is transforming businesses, markets, and value.

By Ashwin Francis

I’ve added a new section to my Claude course, with workflows you can copy and start using right away: Check it out here

What Can Claude Do?

We take Claude through 3 levels of increasingly difficult tasks, which you are likely to see in your workflows.

1. Data Cleaning

Let’s take a dirty dataset and see if Claude is able to save us the effort of reviewing, planning and executing the data cleanup process. I didn’t have a dirty dataset on hand, so I borrowed this one from ForesightBI.

One of the datasets has an issue I have come across many times before: multiple data fields are clumped together in one column instead of being separated by fields.

First, highlight all the affected cells. Then describe the problem that you need to solve. You’ll see the highlighted cells written on top of your prompt.

Before making the edits, Claude will ask you permission to edit the file. This is what the final result looked like.

This example was quite simple because we highlighted the problem and also mentioned what needs to be done.

What if we just highlight the problem and let Claude figure out the best way to clean and represent the data? We prompt:

This data is clumped together. Separate it into a clean table.

The data structure below has multiple product categories with different amounts, all linked to a specific ID. This requires splitting data across columns, transposing the values and aligning identifiers across multiple line items. A bit more complicated.

Highlight all cells in the data structure and tell Claude to ‘Clean up this data’. We had an answer column in the same sheet. I deleted that so that Claude couldn’t cheat. This is the final table it spit out in a separate sheet, which perfectly matches what we want.

Now let’s look at more nuanced workflows. Let’s employ Claude to help us analyze data and find trends.

2. Data Analysis & Business Intelligence

We already saw that Claude is able to help us clean data, but what about summarize, visualize, and synthesize insights?

I downloaded the dataset for this experiment from Kaggle. This was a comprehensive dataset (15,000+ rows) with plenty of fields to support analysis. The dataset tracked Summer Olympic medals between 1976 and 2008. I wanted to conserve my token limits, so I truncated the dataset to cover 2000, 2004, and 2008 (6,000+ rows).

I didn’t know where to start the analysis, so I delegated that to Claude: “Tell me more about this data”.

Claude is smart enough to understand that it doesn’t need to go through the entire dataset to understand the context. Claude ran through the first 10 rows and gave me a summary and data structure. Because it only reviewed the first 10 rows, Claude assumed (based on the title) that the data included 1976- 1996, but this was corrected after we went to the next step.

Next, we need to summarize and visualize. Every executive loves a dashboard, so let’s start by having Claude make one for us:

Some data points that I would like to see: (1) totals medals won by countries, (2) are there any countries that dominate one sport across multiple years, (3) if there are any athletes that have won medals across multiple Olympic events. Additionally we also need visuals to identify trends in the dataset.

This is what I passed to Claude, along with a final line to add any interesting charts/ insights that I might have missed.

The first draft was actually pretty good, but it wasn’t properly formatted with a standardized color theme, chart formatting rules, correct legend positioning, etc. After prompting for some minor tweaks, we got a much better version.

There are still a few edits I could make, like adjusting the row widths and whitespaces between visuals. But this is already a dashboard I can actually use for serious analysis.

The tables answer the three questions that I had. Claude leveraged Python scripts to prepare some insights, as this would not be easy to recreate in Excel manually. This is a huge plus point.

Data analysis helps us uncover what is happening. Let’s see if Claude can use data to model outcomes.

Subscribe now

3. Build a Business Model

I love building financial models and business cases. It helps you get a realistic idea (if you are honest with yourself) of how an initiative will pan out.

For the last use case, we will be using Claude to build a business model for a new startup. Let’s take the example of ScheduleAI, a social media SaaS platform that allows creators to schedule content across multiple platforms, including LinkedIn, Instagram and Substack.

We need a business model to understand what we can realistically achieve in 3 years, and whether the business model is profitable and scalable. We need the model to capture both business metrics like subscriber count, % of paid customers, and churn. Let’s assume that ScheduleAI follows a subscription-based pricing model with 3 tiers, 2 of which are paid.

I brainstormed the business model (pricing tiers, business metrics, etc.) in Gemini and asked for detailed prompts to help Claude build this model in Excel. I find Gemini ideal for brainstorming. They have generous usage tiers, so you can redraw your plans multiple times without worrying about context limits.

Ask Claude to ensure:

1. Values are not hard-coded and linked to primary cells where applicable - model is dynamic

2. Input cells should be formatted differently from formula cells (we follow blue for input and black for formulas) - makes sure that the model you build is dynamic and that you can tweak assumption values to see the impact across the entire model.

Let’s build this out in stages:

Stage 1: Build the assumptions and Subscriber Waterfall

Let’s start with all the assumptions we will use for the model and the baseline business metrics, which we predict for the business. Gemini gave us a detailed prompt with all the key assumptions that fit the ScheduleAI business model

In a blank sheet, create an 'Annual Assumptions' table in Rows 1-15 with columns for Metric, Year 1, Year 2, and Year 3. Set these blue-font inputs: Starting Free Users (5,000), Premium Price ($180/yr), Ultimate Price ($300/yr), Annual Churn (15%), and Free-to-Premium Conversion (5%).

Below this, starting at Row 20, build a 3-year annual subscriber waterfall. All formulas must use absolute references ($B$2 style) to link back to the Assumptions table. Format all hardcoded inputs in Blue Font and all formulas in Black Font. Calculate the annual totals for Free, Premium, and Ultimate users.

Stage 2: Annualized P&L

These assumptions will be the key drivers of the profit and loss statement (P&L) build. The P&L will help us discern if this solution can actually be profitable and that we won’t lose all of our money building and running this into the ground. Gemini gave us a prompt that outlines the P&L build, along with the revenue and cost drivers.

Directly below the User Waterfall, build a '3-Year Annual P&L.' Link Revenue lines to the annual user counts from Stage 1 and the Tier Prices from the Assumptions table. Include rows for COGS ($2.00/user annual hosting) and Operating Expenses (Staff: $50,000/yr; Marketing: 20% of Revenue). Ensure every cell in this section is a formula; no hardcoded numbers are allowed. Apply professional formatting: black font for all formulas and a light green fill for the 'EBITDA' row. Format all currency to $USD with zero decimals.

The P&L is quite high level, we do need to add multiple cost lines here, including incorporation costs, API costs (every time a creator schedules a post on LinkedIn/ Instagram, we will have to call their API), affiliate marketing costs, etc., etc. But this is an iterative process, so we build them out as we identify new revenue/ cost drivers.

3. Model Audit

Now that the first draft model is ready, it’s time to make sure it’s logically sound and paints a realistic picture of the future of ScheduleAI.

We identified some issues in the audit. For example, increasing the headcount didn’t increase staff costs (formula linking issues). Another issue is that the model doesn’t assume user growth over the years—we were doomed to have 5,000 users forever 😅. We should assume the user base grows each year, either at a fixed rate (say 10% annually) or a rate that accelerates over time.

Along with these fixes, I asked Claude to add a legend table so that the model is easy to understand.

The audit step is as important as the build. When you find issues, it’s important to highlight the cells and question Claude on its choices. I spent around 10-15 minutes questioning Claude on the assumptions used and tweaking the model to ensure that dynamism is maintained and the model is logically sound.

Subscribe now

Best Practices for Claude in Excel

If you want professional-grade results, follow this framework:

1. Describe the Mess, Not the Fix: As we did with the Olympic data, don’t tell Claude how to split columns. Just say, “This data is clumped together; separate it into a clean table.” Claude is better at figuring out the logic than you are at explaining it.

2. Request Dashboard Mode: When analyzing large sets, don’t ask for a single chart. Ask Claude to “Create a dashboard in a new sheet with key trends.” This forces it to use Python scripts (as seen in our Olympic test) to do the heavy lifting that Excel formulas can’t handle.

3. Color-Code Your Cells: To avoid the “broken formula” issue we found in the startup model, always prompt: “Format static inputs in blue and formula-linked cells in black.” This makes it easy for you to audit the “thinking” and catch growth rate errors before you present.

4. Always Ask Why: Never trust the first draft. If a headcount increase doesn’t change costs, highlight the area and ask, “Why isn’t this cell linked to the staff cost formula?” The value isn’t in the build; it’s in the audit.

Conclusion: The Shift in Value

The shift we’re seeing in 2026 isn’t just about a new tool. It’s a change in value.

In the past, Excel experts were the only people who knew the most complex formulas and could build a model fast. Today, that person is the one who knows what needs to be done and can validate the output. The new value lies in auditing the logic and interpreting the output.

By offloading the formula-building, the data-scrubbing, and the debugging, you finally get to spend your time on the strategic questions that actually matter. Our time gets invested in making decisions based on data rather than organizing and structuring analyses.

Don’t forget to subscribe to Cash&Cache for AI build guides, product teardowns, and strategies that help you figure out what actually works.

Share

Hi!

I wrote today’s post in collaboration with . Gencay writes LearnAIWithMe, a newsletter focused on practical AI tutorials for people who’d rather replace than be replaced by AI.

If you’re interested in building real projects with AI, I highly recommend checking out LearnAIWithMe

One of my favorite Claude Cowork tasks is web scraping.

Picture this: a bot visits a website for you, reads the page, and pulls out exactly what you need: prices, names, job listings, you name it (so you don't have to copy-paste anything manually).

That’s what web scraping is.

Data is everywhere. Heck, without internet data, tools like ChatGPT and Claude wouldn't even exist. The thing is, web scraping used to be a programmers-only thing.

Not anymore.

Cowork lets you scrape websites by just describing what you want in plain English. In this guide, you'll learn two ways to web scrape:

1. Scrape with Cowork alone: Best for grabbing data from simple pages, but it’s limited to what a single browser session can do and it can struggle with JavaScript-heavy sites

2. Scrape with Cowork + the Apify connector: Best for robust, scalable scraping across complex sites. The only con is that you need monthly credits. That said, Apify's free plan gives you $5 in monthly credits, which is enough for casual use.

By the way, I'm not associated with Apify. I just think it's a solid option worth knowing about. That said, if you're new to web scraping, don't worry about it. Option 1 is a great place to start.

Fun fact: 16,401 people took my web scraping video course back when coding was the only option.

I’ll keeping making guides like this even if they make my own old content obsolete. Why? I believe powerful skills should be available to everyone.

Subscribe for more content like this (& get these perks)

Subscribe now

1. Web scraping with Cowork

For this section, I’ll use a problem I often have as a writer. You might not have the same issue, but this example will give you a good idea of how to web scrape with Cowork.

Every day, I publish Substack Notes, which are short-form posts that help me share bite-sized AI knowledge. Substack doesn’t offer Notes stats or an API to help writers keep track of notes.

Here’s when Claude Cowork comes in.

I had Cowork scrape Substack, build a dataset, and run a monthly analysis.

Here’s what a Note looks like (the data to extract is inside red boxes):

Here are the steps I follow to web scrape with Cowork:

1. Find your target page: Locate the website that has the data you need. In my case, it was the page where Substack archives all my published notes: artificialcorner.com/notes

2. Define the fields you want to extract: Decide exactly what data points matter. I wanted to extract six fields per note: date, post text, number of likes (hearts), number of comments, number of restacks, and post link.

3. Spell out the edge cases: Think about the details that aren’t obvious at first glance — this is what separates clean data from messy data. For example, my notes archive also includes notes I restacked from other writers, which I didn't want. So I told Claude to only extract notes authored by "Frank Andrade" and to limit the pull to a single specific month.

By the way, you will need the Claude in Chrome connector enabled in Cowork: Click “+” → Connectors → Toggle on “Claude in Chrome.” Also, you need the Claude extension active in Chrome (go here and install it).

Here’s the prompt I created from steps 1, 2, and 3:

Use Claude in Chrome for this task.

Start from: https://artificialcorner.com/notes
Author name: Frank Andrade
Month: February 2026

Extract this data from the site keeping in mind the author and month I wrote before: Date, Text of Post, Number of Likes (Hearts), Number of Comments, Number of Restacks, Link of Post

Your first test probably won't be perfect. Review the data Claude collected, and if something looks off, tell it what to adjust for the next run. In my case, the extracted links and a few other data points were wrong.

Here's the feedback I gave Claude

Claude understood everything and returned the right data.

Subscribe now

How to automate your web scraper in Cowork

Once Claude learns the best way to scrape the site, there are two ways to automate this process (so you don’t repeat the prompt and feedback again):

• Turn the chat into a skill: Save your chat as a reusable skill that you can run anytime. In this case, the skill will help me scrape data from any Substack writer’s URL (not only mine)

• Schedule the task: Automate the task for a specific writer on a recurring schedule (e.g., scrape “Frank Andrade“ notes every month on the 14th).

To easily turn a chat into a skill, do this: In the chat, click the dropdown arrow and select “Turn into skill.“ Claude will do the rest.

For this particular case, I used a prompt to guide Claude on how to turn my chat into a skill:

Turn this task into a skill. Keep in mind the fixes we made and that the inputs needed for this skill to work are:

Start from: [link]
Author name: [name]
Period: [month/dates]

Thanks to this, Claude can now scrape not only my notes, but the notes of any Substack writer. I just need to provide their link.

As for scheduling tasks, you can click the dropdown arrow and select “Schedule,” or use the /schedule command. The /schedule command gives you more flexibility — the dropdown only offers daily, weekly, or weekday options, while /schedule lets you set any custom recurring schedule.

In my case, I opened a new chat and used /schedule to have Claude automatically scrape my notes for the previous month on the 14th of each month.

/schedule run this task monthly on the 14th
Start from: http://artificialcorner.com/notes
Author name: Frank Andrade
Period: Previous month

Subscribe now

2. Web scraping with Cowork + Apify connector

When you feel Cowork isn’t enough to scrape a site, use the Apify connector to supercharge your web scraping.

Apify is a web scraping and automation platform with over 23,000 ready-made actors. Each actor is a pre-built scraper for a specific website or data source. Reddit, Google Maps, Meta Ad Library, LinkedIn, Amazon, they are all there, ready to run with no code.

Apify is great for robust, scalable scraping across complex sites. With the Apify connector, Cowork can tap into thousands of specialized scrapers with built-in proxy rotation and anti-bot handling that Cowork alone can't do.

You just pick the actor, set your inputs, and get structured data back in minutes.

Apify Use Cases

This section was written by who writes LearnAIWithMe. It covers how to connect Apify with Cowork and 3 use cases I found very useful:

• Use Case 1: Reddit Trend Tracker

• Use Case 2: Google Maps Business Scraper

• Use Case 3: Meta Ad Library Spy

LearnAIWithMe

If you want to build with AI, not just learn how to prompt, this is the right place.

By Gencay

How to Connect Apify with Cowork

Read more

Every month, there’s a new tool. Every week, there’s a new AI feature. If you’re trying to keep up, it feels like you’re always behind.

Most people deal with this by jumping from tool to tool, tutorial to tutorial, without ever getting good at any of them. They consume a lot but retain very little.

I used to do the same thing.

After teaching AI and other topics on YouTube and Udemy to thousands of people, I realized the problem isn't a lack of resources. It's a lack of structure.

People who learn AI quickly don't watch more tutorials. They follow a system.

Here’s my system for learning Claude and other AI tools without wasting time or feeling overwhelmed.

So far, my guides have taught you what to learn in Claude — the features, the use cases, etc. This one is different. It's about how to learn it.

This guide is exclusive to paid subscribers :)

Every week I publish a free and paid guide. Paid subs get full access to my best guides and other perks.

Subscribe now

Read more

That's what happens when you use Claude without Skills.

Skills is a feature that lets you create reusable instructions so Claude follows your standards — your brand colors, your formatting, your way of doing things (no more generic AI outputs).

Here’s what I got when I tried the same prompt, but using a skill that follows Anthropic's brand guidelines. It nailed the colors, the design, everything!

That’s the difference Skills make.

Just like that, you can use skills to create custom workflows, save hours on repetitive tasks, generate content that follows your company’s brand guidelines, chart styles, font choices, visual design principles, and more.

We’ve searched all over the internet to find the best Claude Skills (so you don’t have to). In this guide, we’ll see:

• How to use Claude Skills

• The best 11 Claude Skills

  • Skills for Presentations & Visuals

  • Skills for Marketers & Business Owners

  • Skills for Business Operations

  • Skills for Building Without Code

• How to quickly install the Claude skills

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

How to use Claude Skills

To work with Skills, first enable this feature.

Go to Settings → Capabilities → Turn on Skills

You don’t need to be technical to create a skill. Claude has a built-in skill creator that builds skills for you. Here’s how to enable the skill creator:

1. Go to preferences in the left sidebar

2. Click “Skills“

3. Select “skill-creator“

4. Toggle skill-creator on (leave it always on)

In the Skills section, you’ll see skills created by Anthropic.

If you toggle on the skill brand-guidelines and try the prompt below, you’ll get slides with the Anthropic colors, design, etc (like the example I showed before).

create a quick presentation on claude skills. use anthropic brand guidelines

There are two ways to use a skill.

You can click on the three dots and click “Try in chat“ or simply let Claude figure out which skill to use. In my experience, Claude is quite good at matching your request to the right skill — just make sure you know what each of your installed skills does so you can phrase your prompts accordingly.

See whether Claude picked the right skill by reading the first lines of the response.

Pro tip: You can turn any Cowork chat into a reusable skill. Just click the dropdown arrow in your chat name and select "Turn into skill."

Subscribe now

The Best Claude Skills

There are many skills out there. Here are the best we found

Skills for Presentations & Visuals

1) Slide Deck Builder

Problem:
Making a presentation takes way longer than it should.

And it’s not the design that slows you down.

It’s everything before that — figuring out what to say, in what order, what gets its own slide, what needs more context. That’s where the time goes. You sit there staring at a blank slide and somehow an hour disappears.

The Skill:
You give it a rough idea: a few sentences, a topic, whatever you have. It turns that into a full slide sequence that already makes sense, applies the visual style you set, and hands you back a real first draft.

Why it works:
It skips the two parts that kill your time: figuring out the structure and making it look like something. Instead of starting from zero, you start by improving something that already has a shape.

Link: Watch the tutorial | Download the skill

2) Explainer Graphic

Problem:
Some topics just don’t click when you read them.

You go through the whole explanation, you get to the end, and somehow... nothing sticks. So you read it again. Then again.

The skill:
You give it a concept — anything that feels hard to explain in words alone. It turns that into an HTML explainer with simple analogies, visual steps, and short sections that are easy to follow.

Why it works:
Switching the format — from a wall of text to a visual walkthrough — is sometimes all it takes. Same concept, completely different experience.

Link: Watch the tutorial | Download the skill

Skills for Marketers & Business Owners

3) Branded Carousel Generator

Problem:
Turning your ideas into social content is one thing. Turning them into content that actually looks like your brand is another.

Most people end up stuck between two bad options: reuse a generic template that doesn’t feel right, or spend way too long tweaking every post trying to make it look consistent.

The Skill:
This one takes your content and turns it into a branded carousel.

Not a generic post split into slides. A carousel shaped around your material, styled in a way that feels like something you’d actually publish — and could publish again next week without starting over.

Link: Download the skill | Download the skill

4) Brand Applicator

Problem:
You finish a document and it’s technically done.

The writing is there. The structure works. But something still feels off. The branding

Wrong font. Colors that don’t match. No logo. It looks like a draft even when it isn’t one.

The Skill:
You hand it your brand rules (fonts, colors, logos) and it applies them to your document automatically.

The output looks consistent with the rest of your business assets. Not like something you threw together before a deadline.

Why it works:
Most people leave this part for the end. And then they rush it, skip it, or send something that doesn’t look as polished as the work inside it.

This handles the last-mile step for you. Not the thinking. Not the writing. Just the part that makes something feel like a real deliverable instead of a working draft.

Link: Watch the tutorial | Read the guide | Download the skill

Subscribe now

5) Market Competitors

Problem:
Competitor research usually ends the same way.

You look at a few pages, notice a few patterns, write something down somewhere — and then move on without ever turning any of it into something you can actually use.

The research happened. The useful output didn’t.

The Skill:
market-competitors takes a website or market input and turns it into a competitive intelligence report.

Not scattered notes. A structured deliverable — the kind you can actually reference when you’re making positioning decisions.

Why it works:
Competitive analysis is only useful if it leads somewhere.

Most of the time it doesn’t because the research never gets a format. It stays in a tab, a doc, or your head until something more urgent takes over.

This gives it a format. And that’s what turns observation into something you can act on.

Link: Watch the tutorial | Download the skill

Skills for Business Operations

6) Workflow Visualizer

The Skill:
Workflow Visualizer turns a process, business system, or operating setup into an interactive visual map.

Phases, triggers, components, connected tools — all in one place, showing how they fit together instead of just listing what they are.

Why it works:
A list of steps tells you what exists. A map tells you how it works.

That difference matters when you need to explain the system to someone else, find where something is breaking, or figure out what to improve without accidentally touching something you shouldn’t.

Link: Watch the tutorial | Download the skill

7) Contract Reviewer

Problem:
Reading a contract quickly is easy. Reading it well is a different thing.

The hard part isn’t the length — it’s knowing what to look for. The clause that sounds fine until something goes wrong. The protection that should be there but isn’t. The term that’s standard until it isn’t.

The Skill:
You upload the contract and it breaks it down into what actually matters.

Key terms, red flags, yellow flags, missing protections, and what you may want to push back on before signing.

Why it works:
It doesn’t replace a lawyer. It does something more useful for most people. It tells you where to pay attention first.

That alone changes how you approach the document — instead of reading every line with equal worry, you know exactly where to slow down (and where to relax).

Link: Watch the tutorial | Download the skill

8) Decision Council

Problem:
Some decisions are too important to look at from one angle only.

The real risk isn’t always making the wrong call. Sometimes it’s making a fast one — without enough pushback, without anyone asking the uncomfortable question, without real tension between perspectives.

The Skill:
Decision Council runs multiple expert personas at the same time, each looking at your decision from a different angle. Then it brings them together into one synthesized recommendation with clear next steps.

You get a structured clash of viewpoints first and then the answer.

Link: Read the guide | Download the skill

Subscribe now

9) Quick Research

Problem:
Research gets messy fast.

You start with one question. That leads to another source, then a different opinion, then three more pages of context you’re not even sure you need.

By the time you’re done, you have a lot of open tabs and no clear picture.

The skill:
You give it a topic and it compresses everything into a structured brief.

I used it to break down a new Claude Cowork release. Summary, features, requirements, pros, cons, pricing, and the actual bottom line. All in one place, already organized.

Link: Watch the tutorial | Download the skill

Skills for Building Without Code

10) Frontend Design

Problem:
A lot of AI-generated interfaces technically work.

They just don’t look like anyone cared.

The code runs fine, but the result feels like something that was assembled in five seconds and never looked at again. Generic layout, default fonts, nothing that feels deliberate.

The Skill:
Frontend Design is an official Claude Code skill that pushes the output toward something more production-ready.

Better typography. Stronger visual decisions. UI that actually looks like someone made a choice — instead of just accepting whatever came out first.

Link: Watch the tutorial | Download the skill

11) Implement Design

Problem:
Going from Figma to code sounds simple until you actually try it.

That’s when the small mismatches start showing up. Spacing is slightly off. Components drift from the original. And the final result looks close — but anyone who worked on the design can tell it isn’t right.

The Skill:
Implement Design is a public Figma skill for translating designs into production-ready code with much tighter visual fidelity.

This skill gives Claude a more structured way to move from design to implementation — so the gap between what was designed and what gets built stays as small as possible.

Link: Watch the tutorial | Download the skill

Subscribe now

How to install the skills

Follow the steps below when you want to add one skill to Claude (.zip or skill.md file)

Customize → Skills → “+” → Create skill → Upload a skill → Upload your .zip or skill.md file

If the .zip file contains multiple skills (as in some of the files shared in this guide), you’ll have to add the .zip to plugins:

Customize → Personal plugins → “+” → Create plugin → Upload plugin → Upload your file

One more thing

You now have 11 skills that can save you hours every week.

But skills work best when you actually know how to use Claude well — the prompting, the workflows, the setup that makes everything click.

That’s what paid subscribers get:

• Free access to all my video courses (over $200 in value)

• A Google Doc with full access to my best Claude guides (updated every week)

• A private community where you get advice from other professionals and from me

I’m adding new courses and guides constantly. Subscribe now and you lock in today’s price for everything that’s already there (and everything I add next!)

What was your favorite skill? Let me know in the comments

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

Hi!

I have a quick question. Please take a second to answer, it’ll help me plan future content :)

Many of you know me from Medium.com. I started writing there in 2021, grew to 149k followers, and used that to build this Substack and two YouTube channels (around 50k subscribers each). So I have a lot of experience in writing and AI.

I’m thinking about sharing these two things with you:

1. My complete Claude writing system: All the files to make Claude write like you in posts, emails, reports, presentations, etc. No guides or tutorials. You just download it and start using it right away.

2. Guides on using AI for writing: How to use my Claude system to write more without sacrificing quality, repurpose content into other formats, etc. Plus everything I’ve learned in these 5 years as a writer to build an audience and make money.

Reply to this email or leave a comment: are you interested in 1, 2, both, or none? I’ll read every single reply. Thanks in advance!

You can think of today’s post like a mix of 1 & 2. Hope you like it!

In 2023, many people were against AI writing.

In 2026, many writers are using AI.

Some writers I follow have admitted they use AI for writing. Others haven’t mentioned it publicly, but they’ve suspiciously increased the number of posts they publish per week over the past few years.

Heck, Substack’s top newsletter in education is called “Write with AI.“

I’m not against AI.

I’m against using AI for bad writing.

If you use AI to enhance your writing and produce high-quality work more frequently, your readers will be happy.

The problem is that most people use AI to produce bad writing.

The fix isn’t a better prompt. It’s a file.

One file that teaches AI exactly how you write, what you never say, and what makes your writing yours. You create it once. You reuse it every time.

In this guide, you will:

• Learn why most produce AI slop

• Get the exact prompt that captures your writing voice in a single file

• Learn why and how to use this file in Claude

• Download my personal writing-voice file as a working example

Subscribe now

Why most produce AI slop

Since ChatGPT’s release, we’ve been obsessed with the em-dash and a few words that make your writing sound AI.

Here’s my (ironic) response to this:

and this

My advice: Don’t mind AI words.

It’s fine if you have a few AI words in your vocabulary.

What’s not fine is to produce content under your name that doesn’t sound like you.

By default, AI doesn’t know your voice. It doesn’t know you hate the word “utilize.” It doesn’t know you always open with a personal story, or that you’d never use a semicolon in a million years.

So it defaults to the safest, most average version of whatever you asked for.

That’s not AI’s fault. You gave it a vague instruction in your prompts and expected a specific result.

The file that fixes this

The fix is an .md file (a simple text file with some formatting). The one we’ll create contains everything Claude needs to write like you.

Not “write well.” Write like you.

Here’s a summary of what goes in it:

• Words and phrases you’d never use

• Sentence patterns you default to

• How you open and close pieces

• Your formatting instincts (short paragraphs? lists?)

• What your writing sounds like at its best

• Writers you admire and what you’d steal from each one

• Positions you’d never take

• and more!

The insight that makes this work: most of a good voice profile is about what you reject. Not “I like direct writing,” but “I’d never use semicolons because they make my writing sound like a college essay.”

That’s the kind of specificity Claude can actually use.

How to create your voice file

This guide is for paid subscribers only. Every week I publish a free and paid guide like this one. To get full access, consider becoming a paid subscriber today 👇

Subscribe now

Read more

For a long time, the way I used Claude Cowork was simple.

Ask for something. Get a result. Move on.

That works great when the task is a one-off, like a quick email, a summary, a draft you need today and forget tomorrow.

But it starts to fall short when the work doesn’t end in a single session. When you need to come back tomorrow or next week, and the AI has no idea who you are or what you were working on. When you spend 10 minutes re-explaining context you already gave.

That’s exactly where Projects changes Claude Cowork.

Projects is a dedicated space where your work lives with its own files, context, and rules ready every time you come back.

In this guide, I’ll show you:

• Why Cowork Projects matters

• How to create your first Cowork project

• How to shape a project that works

• A real use case for projects

Some of you know me from my Udemy courses or YouTube videos. I used to create long-form video content, but AI moves so fast that it stopped making sense — a lesson can go stale in weeks, and updating video is painfully slow.

So I switched. Now I publish (and regularly update) Claude guides every week. Think of them as my courses in written form → Click here to check out my Claude course: 16 lectures and counting!

Every week, paid subscribers get step-by-step Claude guides. No technical skills needed. No ads. Just the stuff that actually matters (+ these perks) 👇

Subscribe now

Why Cowork Projects matters

Before Anthropic released projects, Cowork’s biggest problem was its memory. In the past, by default, every Cowork session started from zero.

Now, with Cowork projects, your work lives in a space that already knows what you’re building, has your files, follows your rules, and remembers what you care about (every time you open it)

Think of it this way: without Projects, every time you come back, you’re briefing a new assistant from scratch. With Projects, you’re walking back into an office where everything is already set up and the assistant already knows the job.

Besides persistent memory, Cowork Projects also offers custom instructions, scheduled tasks, and context. All in one space.

How to create your first Cowork project

If you want to use Projects inside Claude Cowork, the path is pretty straightforward:

Open Claude Desktop → Cowork → Projects → Create your first project

You’ll find 3 options to create a new project.

Here’s when to use each option

• Start from scratch → Use it when you want to create a new space from zero

• Import a project → Use it when you want to import a Claude chat project

• Use an existing folder → If you copied my Cowork system, just select here the “YOUR-WORKSPACE“ folder created

One important detail: importing from an older Claude chat Project doesn't bring over the full chat history. It carries over the instructions, files, and project setup. If an older thread has context you still need, a one-time summary file can bridge that gap.

Subscribe now

How to shape a project that works

Not every task needs a project. A project earns its place when the work:

• Repeats weekly, monthly, or continuously

• Depends on context you don’t want to restate every time

• Pulls from several files or sources

• Tends to produce the same kind of output

That fits more work than you might expect — research, writing, ops, sales, client work, reporting, internal briefs. Anything recurring where context and format matter.

A solid project usually has four parts:

1. A specific purpose: Narrow enough to have one clear job. Weekly competitive tracking, market review, leadership briefing prep, opportunity monitoring.

2. Context that doesn’t repeat: The assumptions, priorities, definitions, and decision criteria Claude needs.

3. Reference materials: The files, links, examples, and sources the project draws from every time.

4. Output logic: What kind of deliverable is expected to produce: a short brief, a memo, an executive update, a prioritized list, etc.

When those four are in place, the project is easy to reopen trust and reuse.

A real use case for Projects

If you’ve already been working with Claude (or read this guide), you probably have folders with useful material inside them. Instead of starting from zero, turn one of those folders into a project.

For this example, let’s say you already have a folder like this:

projects/
└── market-review/
    ├── notes.md
    ├── template/
    ├── references/
    └── drafts/

In this case, I’ll use “use an existing folder” to create the project.

Download the market-review folder here

If the folder already has strong context, you may not need much in the Instructions field. Use that field when you want Claude to follow specific rules — how to format outputs, what to prioritize, what to skip.

Once the folder is connected, you can start working directly on top of it. A simple first prompt could be:

Read the files in this Project and turn the current notes into a decision brief for internal review.

The files don’t move. What’s new is that they now sit inside a workspace that carries context across sessions — instead of reading them cold every time.

That context stays inside the Project. Claude can build on past chats, project knowledge, and outputs from that same workspace without mixing it with unrelated work.

Pro tip: Run the same workflow automatically with scheduled tasks.

Scheduled → Create scheduled task → paste your prompt → choose the frequency → Save

For this example, the prompt would be:

Review the files in this Project, identify relevant competitor, pricing, onboarding, or positioning changes, and produce a concise decision brief for internal review with three sections: what changed, why it matters, and what deserves follow-up.

And that’s it. It’s now scheduled.

You can also schedule a task from an existing project chat.

→ For more Cowork project use cases, download these folders and read this guide to copy my prompts to get started with each project

Where tasks, folders, and skills fit

Before moving on, it’s worth clarifying these concepts

A simple way to see it would be this:

Here’s how each one works:

• A folder stores your files, references, and drafts

• A skill captures a specific way of working or solving a task

• A task executes a one-off action

• A Project is what lets all of that live inside the same context

Final notes

• Cowork Projects is most useful when the work is recurring, context-heavy, and tied to a stable set of files or outputs.

• Cowork projects live on your computer (no cloud sync), so you can’t access a project from multiple computers. If you need a project space built around shared chats, knowledge, and instructions, the older Claude chat projects fit that need better

• Subscribe to this newsletter so you never miss guides like this, delivered straight to your inbox 👇

Subscribe now

This one is mind-blowing!

It’s called “Computer Use.” It lets Claude control your entire computer. This means controlling your keyboard and mouse in any app that you allow.

It’s only available for the macOS Claude app, BUT after testing it since its release, I think Windows users can still benefit from a simplified version of it (keep reading and you’ll understand what I mean).

Claude is releasing updates like crazy lately! I spend a lot of time testing and digging into each one so you don't have to. If it's useful, it's in my guides. If it's not, I save you the time

Click here to read the best Claude guides

If you like this Claude series, support me by becoming a paid subscriber :) 👇

Subscribe now

Computer Use is powerful, but Claude tries to avoid it. Here’s why

After doing dozens of tests, I’ve noticed that Cowork doesn’t always use computer use’s full potential, and there’s a reason for that.

Even if you prompt Cowork “use computer use for this task,“ Claude follows this order to execute tasks:

1. Connectors: If a connector is available (Gmail, Google Drive, etc), Claude uses it. This is the fastest and most reliable path.

2. Claude in Chrome: When there isn’t a connector for the tool you need, Claude will navigate the Chrome browser to work on your task

3. Screen interaction: If options 1 and 2 won’t help, Claude uses computer use to interact directly with your screen: clicking, typing, and navigating your desktop apps.

Claude prioritizes the fastest method (1 and 2) to avoid longer sessions and mistakes.

In my use-cases, you’ll notice that Claude avoids as much as possible using screen interaction (aka computer use). Instead, it uses connectors or Claude in Chrome. Is that bad? Not at all, you’re wasting fewer tokens and getting the job done anyway.

That’s why I think Windows users without the computer use feature can still benefit from the use cases we’ll see later (Claude mostly uses Claude in Chrome).

First, let’s see how to enable Claude’s computer use.

How to enable Computer Use

First, download the Claude desktop app: claude.com/download

Open the app → Select Cowork → Open the sidebar:

1. Click on your name

2. Settings

3. General (inside “desktop app” section)

4. Toggle on “computer use“

How to know you’re actually using computer use

The best way to trigger computer use is to make Cowork work with local files or apps. Here are some ideas:

• Create an analysis using local files and then create a formatted report in Excel

• Navigate apps that don’t have connectors (tools your team uses, macOS apps, etc)

• Fill in a spreadsheet with data from multiple sources, format it, and save it

Here’s a test you can do to see what your computer looks like when it’s controlled using Claude’s computer use.

Go to Cowork and paste in this prompt:

Use computer use for this task:

Open the calculator app to calculate 30x200 and then create an apple note to store the result

This might seem silly, but, believe me, sometimes it seems that this new feature is being used, but it’s not! If you see the orange outline on the whole screen, Claude is controlling your entire computer with computer use. If you see the outlines only in the Chrome window, Claude in Chrome is being used.

These are only technicalities. At the end of the day, we care to get things done without wasting too many tokens. That’s what we’ll do in the following use cases.

Subscribe now

My Best Use Cases

Computer use helped me automate my busywork. Here are my best use cases that you can adapt to your workflow.

#1 Automate LinkedIn responses (“Comment X, Get Y”)

If you use LinkedIn or Instagram, you probably came across the “Comment X and I’ll send you Y.“ X is a keyword that triggers an automation. Instagram has third parties that make this easy. However, as far as I know, LinkedIn doesn’t have any.

With the prompt below, I could automate this:

Use “Computer Use” to control my computer and do this task:

Open the Excel file “LinkedIn DMs.” Find today’s row, copy the message, and open the LinkedIn post. Then open the comments section and send a DM to everyone who commented, using the keyword. After you successfully sent the DM reply “Sent!” to their comment

Why do content creators use this formula? Platforms like LinkedIn penalize posts with links, so this is the best way to share links without hurting views.

By the way, if you want to connect with me on LinkedIn, click here: https://www.linkedin.com/in/thepycoach/

I accept everyone :)

#2 Accept connections

Every day, I get new connection requests on LinkedIn. I manually accept each of them because there’s no native LinkedIn way to automatically accept requests.

Thanks to the prompt below, I can get my connection request inbox down to zero:

use “computer use” to accept all my connection requests on linkedin

Before this test, I was manually accepting a lot of requests. Luckily, a new one came in right when I needed it (lucky me!)

One less task in my daily routine!

#3 Create a customized short link after publishing a new guide

Every time I publish a new guide, I manually create a Bitly shortlink to share it on other platforms. I use the original link slug to create a customized shortlink.

This prompt automated this process:

Use “computer use” for this:

I have a Substack newsletter at artificialcorner.com where I publish AI guides every Wednesday and Sunday. At the end of each Wednesday/Sunday I want you to go to my bitly.com account and create a customized bitly link using the same URL slug used in the original guide (if it’s taken come up with something similar)

Note: This week I published on a Tuesday (it was an exception)

#4 Create an Excel table with web data

To quickly keep track of all the links I create with Bitly, I need an Excel table with these columns: Guide name, Original Link, Bitly Link.

The prompt below extracts the data I need from Bitly and puts it into Excel.

Use “Computer Use” for this:

Build a table in Excel with these columns: Guide name, Original Link, Bitly Link. All the info is in bitly. Only consider the links from my site artificialcorner.com

#5 Update an Excel table

If I need to update a table with fresh data, I ask Claude to visit a website, extract the information I need, and update the table for me.

Here’s the prompt I use:

Use “computer use” for this:

Go to bitly.com and add the latest link created to my dataset Bitly_Links_ArtificialCorner.xlsx located in Download -> Bitly. Open the Excel file and add the data in a new row. The data needed is Guide Name, Original Link and Bitly Link

Final tips to get the most out of computer use

• You can turn a chat into a scheduled task, so it runs on any day/hour you want. Learn how here: artificialcorner.com/p/claude-automates

• You can use Dispatch to trigger computer use from your phone. Check out my guide: artificialcorner.com/p/cowork-dispatch

• To prevent Claude from using some apps, do this: Settings → General → Denied apps → Add app

• Close anything you don’t want Claude to see. When Claude uses your computer, it takes screenshots to navigate your screen

• Subscribe to this newsletter so you never miss guides like this, delivered straight to your inbox 👇

Subscribe now

Click here to read my best Claude guides

The tool wasn’t broken. My workspace was.

Most people open Cowork and do this: open a folder → write a prompt → hope for the best

Sometimes it works. A lot of times, it doesn’t.

And that’s not Cowork’s fault. Cowork is stepping into your workspace with zero context, no rules, and no clear way of working.

So it starts improvising.

It doesn’t know what tone to use.
It doesn’t know what it should never assume.
It doesn’t know what files it must always read.

That’s why so many people try it and walk away thinking it feels like just another chat … just with file access.

In this guide, we’ll fix that. I’ll show you a simple system that gets the most out of Cowork.

By the end of this guide, you’ll have:

• A system that gives Cowork the context it needs to stop improvising

• How to use this system with 3 real-world examples

• How to prompt this system (spoiler: you’ll need short prompts — the system does the heavy lifting)

No coding. No complex setup. Just a structure that works.

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

Cowork doesn’t need more prompts. It needs a system

If you come from the chat world, this makes sense: write a better prompt, get a better result.

In Claude Cowork, that is no longer the main piece. The workspace is.

Because Cowork does not just work with what you type into a text box. It works with what it can read inside the folder you give it access to:

- your project files
- your working rules
- your context
- your preferred way of communicating
- your prior drafts
- your examples of success

You need to turn the way you work into a small, reusable system.

Here’s an overview of the system we’ll build:

That jump looks small. It isn’t.

Because the moment your way of working lives in files, Cowork stops starting from zero every time. When that happens, the results stop feeling generic and start feeling like “this thing already knows how I work.”

Quick Cowork Recap: Installation

To use Cowork, you need to download Claude for desktop:

Go to claude.com/downloads → download Claude Desktop → install it → sign in

Once it's installed, select Cowork. That's it. You're in.

Quick setup: Add Global Instructions

Global instructions is where you set the default behavior Cowork should bring into every session. Do this:

Open profile menu → Settings → Cowork → Add your Global Instructions → Save

Copy these global instructions:

You are my working partner inside this workspace.

Start by reading:
1. context/HOW-I-WORK.md
2. system/SYSTEM-RULES.md
3. the relevant folder inside projects/

Do not execute immediately if the goal, audience, or output format is unclear.
Ask questions first when missing context could weaken the result.

Do not invent facts, decisions, dates, numbers, or approvals.
Make uncertainty visible.

Use the workspace files as the main source of truth.
Save final outputs in outputs/.

This gives Cowork a reliable starting behavior before the real context comes from the workspace.

The simple system I’d actually start with

I wanted to strip this down as much as possible.

I wouldn’t start with ten folders.
I wouldn’t start with an entire empty office.
I’d start with four. That’s it.

YOUR-WORKSPACE/
├── system/
├── context/
├── projects/
└── outputs/

On your computer, it looks like this:

That’s already enough to get started.

• system/ → stores the base rules (stays fairly stable)

• context/ → stores the stable context (stays fairly stable)

• projects/ → stores the active work (it changes by project)

• outputs/ → stores what Cowork produces

A small structure that can grow without turning into a mess.

But a folder by itself doesn’t tell Cowork much. What makes the workspace useful are the files that explain Cowork how to work.

The files that bring the structure to life

Folders give Cowork a place to work. Files tell it how to work.

That’s why I’d start with just two:

1. context/HOW-I-WORK.md

2. system/SYSTEM-RULES.md

HOW-I-WORK.md captures who you are inside this workspace: how you work, how you think, how you communicate, and what kind of work you consider good.

SYSTEM-RULES.md defines how Cowork should behave: when it should ask before acting, what it shouldn’t assume, how to handle ambiguity, and what a strong result looks like.

What should live inside HOW-I-WORK.md? A simple version of this file could include:

• what kind of work you do

• who you do it for

• how you think

• how you communicate

• what feels natural in your writing

• what you dislike

• which sources carry more weight when there’s a conflict

• real examples when useful

What should live inside SYSTEM-RULES.md? A simple version of this file could include:

• core operating principles

• when Cowork should ask before acting

• how it should behave during execution

• how to handle ambiguity

• what counts as a strong output

• clear boundaries

• a final review before delivery

The structure would look like this:

Up to this point, everything sounds simple.

Here’s the most important part: I wouldn’t write those files from scratch. Because the moment you do, you start guessing.

You describe the version of yourself you wish were true — not the version that actually shows up in day-to-day work.

That’s why I prefer a different approach: build both files through an interview.

A good interview helps extract things most people wouldn’t write well on their own:

• how they actually work

• how they actually write

• what they value in a strong deliverable

• what makes them distrust a weak one

• what Cowork should never assume

• what kind of help is genuinely useful to them

Then that interview generates the two files that support the entire system.

Copy the prompt below to generate HOW-I-WORK.md and SYSTEM-RULES.md

You’ll be asked about 50 questions, one at a time. Answer each one honestly and in detail for the best results.

You are my workspace interviewer.

Your job is to extract enough high-quality signal to help build 2 things for Claude Cowork:

1. context/HOW-I-WORK.md
2. system/SYSTEM-RULES.md

Your goal is not to interview me forever.
Your goal is to gather enough specific, real, usable information to generate those files well.

## Interview objective

You are helping define how Claude Cowork should work with me inside a reusable workspace.

That means you need to understand:

- what kind of work I do
- what I usually create
- who I create it for
- how I think
- how I write
- what I value in strong outputs
- what I dislike in weak outputs
- how I want Claude Cowork to behave
- what Claude should never assume
- when Claude should ask questions before executing

## Interview structure

You will conduct the interview in 3 blocks:

### Block A - Work context

Use this block to understand:

- what I do
- what kinds of deliverables I create
- who my audience is
- what those people value
- what sources matter most
- what this workspace should help me do

### Block B - Voice, style, and thinking

Use this block to understand:

- how I sound at my best
- how I sound when I am writing lazily or on autopilot
- the words, phrases, and structures that feel natural to me
- what I hate in writing
- how I organize ideas
- how I think through problems
- what I tend to notice, distrust, or emphasize
- examples of real writing that sound like me

### Block C - System behavior and rules

Use this block to understand:

- how I want Claude Cowork to behave
- when Claude should ask before acting
- what makes an output strong or weak for me
- what should never be assumed, invented, or finalized too quickly
- how uncertainty should be handled
- how conflicting information should be handled
- what kind of help is actually useful to me

## Interview rules

1. Ask one question at a time.  
Wait for my response before moving on.

2. Do not ask a fixed number of questions.  
Ask only as many as needed to build the files well, up to a maximum of 50 questions.

3. Push back on vague answers.  
If I say something generic, ask for a specific example.

4. Ask for real evidence.  
Whenever useful, ask for:
- a real paragraph I wrote
- a real sentence I like
- a real sentence I would never publish
- a real example of a good or bad output
- a real example of a decision, document, or deliverable

5. Call out contradictions.  
If I say something that conflicts with an earlier answer, point it out and ask me to resolve it.

6. Distinguish reality from aspiration.  
If I describe the person I wish I were instead of how I actually work, write, or decide, call it out and ask for proof.

7. Go deeper only when useful.  
If an answer contains strong signal, follow it.  
If not, move on.

8. Do not flatter me.  
Do not say “great point,” “interesting,” or similar filler.  
Just ask the next question or push deeper.

9. Optimize for usable specificity.  
Do not collect abstract opinions if they will not help build the files.

10. Stop when you have enough signal.
Do not drag the interview longer than needed. Stop earlier if you already have enough strong signal to generate both files well.

## What you are building toward

At the end of the interview, you will generate 2 outputs:

### Output 1

context/HOW-I-WORK.md

This file should include:

- what kind of work I do
- what I create
- who I create for
- what my audience values
- how I think
- how I communicate
- what feels like me
- what I dislike
- what sources carry more weight
- real examples when useful

### Output 2

system/SYSTEM-RULES.md

This file should include:

- core operating principles
- when Claude should ask before executing
- how Claude should behave during execution
- how to handle ambiguity
- how to handle conflicting information
- what a strong output looks like
- what weak outputs get wrong
- collaboration preferences
- hard boundaries
- a final pre-delivery check

## Output requirements after the interview ends

When you judge that the interview is complete, generate both outputs in this exact format:

[context/HOW-I-WORK.md]
...

[system/SYSTEM-RULES.md]
...

After that, add this section only if needed:

[Missing or weak signal]
...

Do not generate the files before the interview is complete.

## How to begin

Start with Block A.
Ask me the first question.

Why does this approach actually work?

You’re extracting real info to build two files that will be reused again and again.

That’s the shift that matters.

No more files just for the sake of having more files. No more long prompts to compensate for missing context. Just a minimal structure, two well-built files, and a smarter way to give Cowork the context it needs.

Subscribe now

How to apply this method: 3 real-world examples

The advantage of this structure is that it doesn’t depend on a single task.

It works for an article, report, presentation, research project, weekly brief, proposal, document synthesis, etc

The logic is always the same:

project folder + 2 core files + short prompt

That’s the pattern.

You don’t need to reinvent how you work for every task. You only need to change the content inside projects/ and give Cowork a short prompt to trigger it.

In other words:

• system/SYSTEM-RULES.md stays fairly stable

• context/HOW-I-WORK.md stays fairly stable

• what changes each time lives inside projects/

The projects folder is where the specific work lives — the brief, notes, references, drafts, and any file tied to that particular task.

Click here to download all the folders and files used in the examples below

Example 1: Turn scattered notes into a decision brief

Inside projects/, you create this (it may vary by project):

projects/
└── market-review/
 ├── notes.md
 └── template/
 ├── references/
 └── drafts/

The system doesn’t change. The only new thing is the content inside market-review/

Notes.md could look like this:

# Market Review Notes

## Context
We are reviewing competitor moves from the last 30 days.

## Raw observations
- Competitor A launched a lower-priced offer
- Competitor B simplified onboarding
- Competitor C is publishing more case studies

## Risks noticed
- We may overreact to pricing changes
- We do not know yet if these moves are converting better

## What I need
Turn this into a decision brief for internal review.

references/ → previous presentations, competitor reports, benchmarks, internal documents, anything that helps interpret the notes

drafts/ → first version of the brief, a condensed version, a more executive version if you need one

The prompt could be as simple as:

Read the project folder first.
Then ask me questions before you execute.
I want a decision brief, not a generic summary.
Save the result in outputs/briefs/.

The method didn't change. Only the content did.

Here are the results:

The brief grounds the decision, separates risks and options clearly, and ends with an actionable recommendation and concrete next steps.

Example 2: Build an executive report for a meeting

Now consider this case:

projects/
└── q3-review/
 ├── brief.md
 ├── notes.md
 └── template/
        ├── references/
        └── drafts/

You're not working with scattered notes anymore. Now you have material for an executive report.

brief.md could look like this:

# Q3 Review Brief

## Goal
Prepare an executive report for the leadership meeting.

## Audience
Senior leadership team.

## What matters most
- What changed this quarter
- What risks need attention
- What decisions may be needed next

## Constraints
- Keep it concise
- No speculation without labeling it
- Focus on implications, not just activity

And notes.md could look like this:

# Q3 Review Notes

## Highlights
- Revenue held steady
- Customer retention improved slightly
- Two initiatives are behind schedule

## Open questions
- Whether resources need to be rebalanced
- Whether delays are operational or strategic

references/ → metrics, previous quarterly reports, financial documents, team summaries, older slides

drafts/ → first executive report, a shorter version, a risk-focused version, a structure ready to turn into a presentation

The prompt:

Go through the project folder first.
Use my quality standard and workspace context.
Ask me questions before you write.
Then create an executive report for review and save it in outputs/reports/

Here are the results:

Cowork turned scattered context into a clear executive report with metrics, risks, and decisions ready for discussion.

Example 3: Draft a presentation without starting from zero

The folder could look like this:

projects/
└── strategy-update/
    ├── brief.md
    └── template/
        ├── references/
        └── drafts/

And brief.md could look like this:

# Strategy Update Brief

## Goal
Create a presentation outline for a strategy update meeting.

## Audience
Internal leadership and cross-functional leads.

## What the deck should do
- explain what changed
- show why it matters
- align people on next steps

## What I don't want
- slide titles with no message
- too much text
- generic consulting language

references/ → metrics, previous quarterly reports, financial documents, team summaries, older slides

drafts/ → first structure, a condensed version, a more risk-focused variant, an outline ready to turn into slides

The prompt:

Read the project folder first.
Then ask me questions before you execute.
I want a presentation outline with a strong flow, clear sections, and what each slide should do.
Save it in outputs/drafts/

Here are the results:

Cowork didn't just organize ideas. It turned a fuzzy strategic change into a presentation with a clear narrative.

How to prompt Cowork using this system

At this point, you already have everything you need to make Cowork work.

But there’s an important difference between a tidy workspace and one that actually saves you time: how you use it once it’s already up and running.

Here are my two favorite ways:

#1 Question-first flow

A question-first flow is just a way of using Cowork.

The logic:

• Cowork reads the folder

• then it asks you questions if context is missing

• only after that does it build the deliverable

This is activated by how you give the instruction.

For example:

Read the project folder first.
Then ask me questions before you execute.

You already saw this in the examples:

• In the first one, Cowork asked what the real decision was

• In the second one, it asked about the initiatives and the decision that should come out of the meeting

• In the third one, it asked about the workstreams, the next steps, and the format

That was question-first flow in action.

It makes sense to use it when:

• the goal isn’t fully clear yet

• the audience matters

• the result has real impact

• one bad assumption could ruin the deliverable

• there are several possible outputs and you want Cowork to pick the right one

In other words, use it when you don’t want Cowork filling in the gaps on its own.

#2 Short prompts

A short prompt is simply a prompt that works well because the workspace already has context.

That point matters.

In an empty chat, a short prompt usually falls short. But in Cowork, if you already have:

system/
context/
projects/
outputs/

...the prompt no longer has to carry all the weight.

The context lives in the files. The prompt just activates the system.

For example:

Read the project folder first.
Then ask me questions before you execute.
Create the deliverable and save it in outputs/.

or

Use my HOW-I-WORK and SYSTEM-RULES files.
Then use the project folder to create a first draft in outputs/.

Short. It’s not magic. Your workspace is already doing the real work.

If you need to write a very long prompt just for Cowork to understand the basics, the problem usually isn’t the prompt. The problem is that your workspace isn’t carrying enough context yet.

How to grow the system without turning it into a mess

Once people understand this concept, they usually make one of two mistakes:

• They stick with a structure that’s too thin

• or they try to build a huge system from day one

You don’t want either one.

Here’s the core idea: start small and let the work tell you what’s missing.

Grow only when a pattern repeats.

Don’t create files because they might be useful someday. Create them when you notice something is already repeating and deserves its own file.

When it makes sense to expand system/

• Cowork keeps making the same mistake over and over

• you have rules that apply almost all the time

• your quality standard is clearer now

• you want to lock in a way of working that shouldn’t change from project to project

For example: if you notice Cowork keeps confusing a summary with a recommendation, that already deserves a more explicit rule inside SYSTEM-RULES.md or even a small new section.

When it makes sense to expand context/

• your audience already deserves its own file

• your writing style needs more precision

• you want to separate references, tone, or source-of-truth rules

• your way of working no longer fits inside a single file

The goal is context that’s clear and actually usable.

When it makes sense to expand projects/

• a certain kind of work always uses the same files

• a project needs more internal order

• you want to better separate the brief, notes, references, and drafts

• a recurring task already has its own logic

This is where most of the change happens anyway.

How to tell if the system is growing well

You’ll notice it because each new task needs:

• less explanation

• fewer corrections

• fewer restarts

• and delivers something more useful on the first pass

If that’s happening, the system is growing well. If not, you’re just accumulating structure.

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

But there was always one problem: You had to be sitting at your desk to use it.

If you wanted to start a task or monitor Cowork’s output while you were at the gym or grabbing coffee, you couldn’t do anything about it. There wasn’t mobile access.

That just changed.

Cowork just got a new feature called Dispatch. This feature lets you text Claude a task from your phone, and it completes the work on your computer while you’re away. When you come back home, the work is done.

In this guide, you’ll learn:

• How to set up Dispatch on your phone and desktop

• How to use Dispatch

• How to unlock Dispatch full potential with skills, tasks, and more

• A few things to keep in mind before you use this feature

I’ve been getting a lot of sponsorship offers for this newsletter lately.

I appreciate every single one, but out of respect for my paid subscribers, I’m keeping this space ad-free.

Click here to read my best Claude guides

If you like my Claude guides and want me to keep making more, consider becoming a paid subscriber👇

Subscribe now

How to set up Dispatch

Here’s what you need:

• The latest version of the Claude desktop app (Mac or Windows)

• The latest version of the Claude mobile app (iPhone or Android)

• A Claude paid plan

Here’s how to set it up:

On your phone:

1. Open the Claude app → open the sidebar → Dispatch → Pair with your desktop

On your computer:

1. Open the Claude desktop app → open the sidebar → Dispatch. You’ll see a screen that says “Coworking on the go.” Click “Get started.”

2. Click “I’m signed in on my phone.”

3. Toggle on two things: give Claude access to your files, and keep your computer awake while Dispatch is running (if your computer goes to sleep, Claude stops working)

The setup is finished!

You should see the screen below when you go to Dispatch in the desktop app (in the mobile app, you’ll see a simplified version)

You now have one continuous conversation with Claude Cowork that works from both your phone and your desktop. Text it a task from your phone, and Claude works on your computer using your local files, skills, connectors, and more.

Let’s see how to use Dispatch with a basic example.

Subscribe now

How to use Dispatch

Dispatch is designed for moments when your phone is the only device you have. Here’s how to fire it up on your phone.

Go to the Claude phone app → open the sidebar → Dispatch

You’ll see this default message on your phone.

Use a prompt to point it to a folder on your computer and tell it what to do.

You can type something like this:

“In my [folder name], can you give me a quick summary of what’s in the final report spreadsheet?”

In your prompt, be specific about the folder and the file name.

In a few minutes, the report summary will be generated on your phone. No laptop needed. Just a text message to Claude and a clean summary back on the phone.

But that’s only the most basic thing you can do with Dispatch!

Now that you've seen it in action on your phone, let me show you the computer setup that unlocks its full potential.

Unlock Dispatch's full potential with skills, tasks, and more

If you’ve used Cowork before, you probably know about skills and scheduled tasks.

Skills are reusable instructions you save inside a folder so Claude knows exactly how to handle specific tasks. Scheduled tasks are prompts that run automatically at a certain hour and day.

Both make Dispatch really powerful, because you can trigger them from your phone with a short message.

Step 1: Add skills to your list

To see the list of skills you can trigger from your phone, do this on the desktop app:

Open the sidebar → Go to Customize → Skills → List of skills available

To download skills and learn how to create your own, read this: artificialcorner.com/p/claude

Step 2: Add tasks to Scheduled

To see the list of tasks you can trigger from your phone, do this on the desktop app:

Open the sidebar → Go to Scheduled → List of tasks available

To learn how to create scheduled tasks, read this: artificialcorner.com/claude-automates

Step 3: Trigger skills or tasks from your phone

Now that you know the list of skills and tasks available, you can trigger them with a short message from your phone.

Here are the steps to follow:

1. Open the Claude phone app → open the sidebar → Dispatch

2. Type “Run the [name of skill/task]“

3. Allow permission

4. You get your output

If your task/skill generates a file, you might need to explicitly tell Claude to show you the file in the phone Dispatch chat.

Subscribe now

Here’s an example

Say I’m in the barber shop or sitting in the taxi. I can grab my phone, open Claude, and type: Run the task “export top paid newsletters to CSV.” That’s a task that scrapes Substack newsletter rankings from a site.

Pro tip: Use simple names for your skills/tasks so they're easy to trigger on your phone.

Done! The CSV file was generated. I can open it on both my phone and computer.

When I come back home and open Claude Cowork on my computer, the same chat is right there in the Dispatch tab.

But there’s one thing in the desktop app that you won’t find in the mobile app — in recent chats, you’ll see one chat tagged “Dispatch.“ That’s a chat that is created every time we run a task using Dispatch (but it’s only visible in the desktop app)

Inside the chat tagged Dispatch, there’s the hidden prompt and process used to carry out my task.

The complete chat is like the longer version of the Dispatch tab. There, I can also find the CSV file generated.

Final Notes

Here are a few things worth keeping in mind:

• Your computer has to stay on: This is the most important thing. If your desktop goes to sleep or you close the Claude app, Dispatch stops working. Toggle on the “prevent sleep” option during setup.

• It’s one continuous thread: There’s no way to start a new conversation or manage multiple threads in Dispatch. Everything lives in a single chat, which means Claude remembers context from previous tasks.

• You can run multiple tasks at once: Send three or five different tasks from your phone and Claude will work on them in parallel (as long as the tasks don’t depend on each other).

• Sequential tasks aren’t supported yet: If task B depends on the output of task A (like “research X, then email the team about it”), you’ll need to send them one at a time and wait for the first one to finish.

• No notifications yet. Claude won’t ping you when a task is done. You’ll need to check back in the chat to see results. Hopefully, this gets added soon.

• Connectors extend to mobile automatically. Whatever apps you’ve connected in Cowork (Gmail, Slack, Google Drive, etc.) are available through Dispatch

Dispatch is still in preview mode, so expect some rough edges. But even in its current state, it’s already changing how I use Cowork.

Subscribe now

Click here to read my best Claude guides

I have too.

But here’s the thing: Claude doesn’t count messages. It counts tokens. Some chats eat through your limit 10x faster than others.

So I put together the best practices to stretch your Claude plan as far as it can go.

You don't have to follow all of them. Think of it like one of those iPhone tutorials to save battery. You could follow every tip, but you wouldn't have as much fun with it if you did.

Pick the ones that fit how you work. Even 2–3 of these will make a noticeable difference.

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

#1 Edit your prompt. Don’t send a follow-up

When Claude’s answer misses the mark, most people send another message like “no, I meant...” or “can you try again but...”

Every follow-up message adds to the conversation history.

Claude re-reads the entire conversation every single turn. So by message 10, you’re not just paying for message 10 — you’re paying for messages 1 through 9 all over again.

This adds up fast.

Instead of sending a new message, click the edit icon on your original message, fix the prompt, and regenerate. The old exchange gets replaced instead of added.

Here’s how to do it

Go to the prompt you wrote → click the pencil icon → rewrite your prompt → hit send. Claude regenerates as if the bad response never happened.

#2 Start fresh every 15–20 messages

Again, Claude re-reads your entire conversation history every single turn.

Your first message costs hundreds of tokens. By message 20, a simple question can cost thousands of tokens. The longer the chat, the more expensive every message becomes.

Long chats are expensive chats.

Start a new conversation every 15–20 messages.

If you need context from the old chat, ask Claude for a summary, copy it, and paste it into the new chat as your first message.

Here’s how to do it

When a conversation starts getting long → ask Claude: “Summarize everything we’ve discussed so far” → copy the summary → open a new chat → paste it as your first message → start fresh

You get fresh context without the bloated token cost.

Subscribe now

#3 Combine multiple questions into one

Sending three separate messages means Claude loads the conversation context three separate times.

Batch your questions into a single message.

Instead of sending “Summarize this article,” then “List the main points as bullets,” then “Suggest a headline,” send all three at once.

Write one message like: “Summarize this article, list the main points as bullets, then suggest a headline.”

One message. Three answers. One context load.

By the way, the answers are often better when Claude sees the full picture.

#4 Upload recurring files to projects

If you’re uploading the same file in multiple chats, Claude is re-counting those tokens every single time.

The solution? Use Projects. Upload your files there once, and they get cached so they don’t re-cost you on each conversation.

This is a huge saver for anyone who works with long documents regularly.

Here’s how to do it

Go to the sidebar → click Projects → create a new Project → upload your recurring files there. Now every conversation inside that Project can reference those files without re-paying the token cost.

Upload once. Stop wasting tokens every time.

#5 Set up memory

Every conversation you start without context burns setup messages just re-explaining who you are and how you work.

“I’m a marketer, I write in a casual tone, I prefer short paragraphs...”

You’re spending tokens on the same intro over and over.

A simple way to fix this is to set up memory:

Settings → Memory → add things like your job, your writing style, your preferences.

A better approach is to create an .md file with context about you.

Read this guide to generate .md files with context: artificialcorner.com/p/md-file

Create it once. It runs forever.

#6 Turn off features you’re not using

If you’re not using web search, research mode, connectors, and others, turn them off

They add tokens to responses even when you don’t need them.

Same goes for extended thinking. Leave it off by default and only switch it on when you want Claude to think longer for better responses.

Here’s how to turn them off

Click + → Turn off research, web search, or connectors

Click model selector → Turn off extended thinking

#7 Use Haiku for simple tasks

Most people use the same model for everything — whether they’re brainstorming ideas or doing deep research.

The bigger the model, the more tokens it costs per message.

Using the right model is the single highest-impact decision you can make.

Use the right model for the job:

• Haiku (very low cost) → Quick answers

• Sonnet (medium cost) → Everyday tasks

• Opus (high cost) → Complex tasks

Here’s how to switch models

Just click the model selector at the top of the chat and pick the right one for the task.

#8 Spread your work across the day

I used to sit down at 9 and power through everything in one session. By 11, I’d already hit Claude’s limit. Done for the day.

What I didn’t realize: Claude’s limit runs on a rolling 5-hour window. It doesn’t reset at midnight. Instead, your usage gradually falls off — messages you sent at 9 AM stop counting against you by 2 PM. So if you burn through everything in one burst, you’re leaving most of the day’s capacity on the table.

Split your work into 2–3 sessions per day, instead of 1.

Here’s what I do.

I do a session in the morning, another after lunch, and another in the evening. By the time I come back for each session, my earlier usage has already aged out of the window, and I’ve got a fresh allowance to work with.

Space them out so your rolling window resets between sessions.

Claude counts tokens, not messages. Now you know how to make every token count.

Some of these best practices were taken from Claude Support. For more, read this

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

You chat with it, write some prompts, use the default model, and wonder why your outputs feel no different from ChatGPT.

Why? You haven’t set up Claude properly.

In this guide, we’ll fix that. You’ll learn how to:

• Set up Claude chat (the features that separate it from ChatGPT)

• Set up Claude Cowork (the files you need before every session)

• Set up Claude Code (for non-technical folks)

New to Claude? Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

How to set up Claude chat

“Claude chat” is the default way to interact with Claude, whether you use the web or desktop app. Here’s how to get the most out of it.

#1 Models + extended thinking

Claude gives you three models to choose from. Think of them as three employees with different strengths.

Opus 4.6: It’s the smartest. It takes its time, thinks deeper, and gives you the most thorough answers.

• When to use it: For complex tasks (long documents, strategy, analysis, etc)

• Downside: It’s slower and eats through your usage limits faster.

Sonnet 4.6 (default model): It’s fast and smart

• When to use it: For quick writing, brainstorming, summarizing, and daily work

• Downside: It’s not as thorough as Opus

Haiku 4.5: It replies almost instantly

• When to use it: For quick questions or simple tasks (Claude won’t think too hard)

• Downside: It’s the least powerful of the three

To switch models, click the model selector.

To be completely transparent, I use Opus 95% of the time.

When I need AI for casual tasks, I use ChatGPT or Gemini instead of wasting usage on Sonnet or Haiku. If you have access to either, I’d recommend doing the same. If not, use Opus wisely.

I also prefer to keep “extended thinking” turned on. When it’s on, you’re telling Claude: “Don’t rush. Think this through before you answer.”Leave it off for everyday stuff (quick questions, casual writing, simple requests).

#2 Projects

Think of a project as a folder where Claude already knows the context. You upload documents, add instructions, and from that point on, every conversation inside that project starts with context.

Here's how to create a project:

1. Click “Projects” in the left sidebar

2. Click “New Project”

3. Give it a name

4. Add custom instructions and upload files (PDFs, docs, etc)

Pro tip: Create an .md file with context about you or your work and upload it to your project, so you don’t have to repeat it over and over again.

📚 How to generate an .md file with context: artificialcorner.com/p/md-file

Subscribe now

#3 Skills

A skill is a set of instructions you save once, and Claude loads it automatically whenever it's relevant.

Say you write weekly reports and they always follow the same structure. Instead of explaining that structure every time, you create a skill for it. Next time you prompt “write my weekly report,” Claude recognizes the task, loads your skill, and follows your format from the start.

To work with Skills, enable it in settings.

Go to Settings → Capabilities → Turn on Skills

You don't need to be technical to create a skill. Claude has a built-in skill creator that builds the skill for you. Here’s how to enable the skill creator:

1. Go to preferences in the left sidebar

2. Click “Skills“

3. Select “skill-creator“

4. Toggle skill-creator on (leave it always on)

If you toggle on the skill brand-guidelines and try the prompt below, you’ll get slides with the Anthropic colors, design, etc.

create a quick presentation on claude skills. use anthropic brand guidelines

📚 Learn more about skills: artificialcorner.com/p/claude

#4 Artifacts

Artifacts turn ideas into shareable apps, tools, or content. Common examples of artifacts are diagrams, flowcharts, websites, and documents.

To trigger an artifact, just describe what you want (an app, diagram, document, or chart) and Claude will automatically create an artifact for it

To turn artifacts on, do this:

1. Click your initials or name in the lower left corner

2. Navigate to Settings > Capabilities

3. Find Artifacts and toggle it on or off

One video is worth a thousand words. Watch artifacts in action:

How to set up Claude Cowork

Claude chat is like texting a smart friend. You ask, it answers.

Cowork is like hiring an assistant. You describe the task, point it to your files, and it gets to work — planning steps, reading documents, creating files, and delivering results directly to a folder on your computer.

Cowork lives inside the Claude desktop app. Here’s how to set it up:

1. Download the Claude app → claude.com/download

2. Open the app and sign in

3. At the top of the window, you’ll see three tabs. Click Cowork

4. Create or pick a folder on your computer where Claude will work

5. Point Claude to that folder

   

6. Try this prompt:

You are going to help me create a context file I can use with Claude Cowork.

Ask me questions one at a time to understand my work: who I am, what I do, how I like things done, what I never want to see in outputs, and what “good work” means to me specifically. Use AskUserQuestion for it.

Ask me 40-50 questions total. One at a time. Push back if my answer is vague — ask for a specific example. When we’re done, compile everything into a .md file I can save and use with Claude.

You’ll get an .md file with context about you. Put it inside the folder Cowork has access to.

Next time you use Cowork, add this at the end of your first prompt:

[your prompt]

Before you start, read [your filename].md first

Now you don’t need to repeat context to Cowork.

Pro tip: To get the most out of Cowork, it helps to keep these three .md files ready:

• about-me.md — everything about you, your work, etc

• voice-and-style.md — your tone preferences, formatting rules, words you like/dislike, writing samples, etc

• working-rules.md — how Claude should behave

📚 How to generate these .md files: artificialcorner.com/p/md-file

Cowork offers something that you won’t find in Claude chat: plugins. They’re pre-built instruction sets that make Claude good at specific professional tasks.

Here’s how to install and use plugins:

Cowork → Click + → Add plugins → Pick the plugin you like → Click install

📚 Learn more about plugins: artificialcorner.com/p/cowork

Subscribe now

How to set up Claude Code

Claude Code is Cowork on steroids.

Both can read your files and execute tasks. But Claude Code goes deeper. It can edit code, run commands, manage entire projects, and work across multiple files at once. It’s the most powerful way to use Claude.

Now, here's the part that scares most people: Claude Code was originally built for the terminal (that black screen with the blinking cursor that feels like hacking in a movie)

The good news is that after the installation, you don’t have to touch the terminal!

Here’s how to install Claude Code:

1. Go to the Claude website → “Install the latest version“ → Look for your operating system → copy and paste the installation command into your terminal → execute it

If you’re new to the terminal, here’s how to open it:

• Windows: press the Windows key, type “CMD” and press enter

• Mac: press Command (⌘) + Spacebar, type “Terminal“ and press enter

After running the command, you should see this message: “✅ Installation complete!“

To verify that Claude was installed successfully, type “claude” and press enter. You should see the image below in your terminal

2. Install VS Code. It has a friendlier interface to work with Claude Code

3. Install the Claude extension within Visual Studio Code: Open VS Code → View → Extensions → search for Claude → click Install

4. After installing the extension, click on the Claude icon to log in to Claude

5. Finally, open a folder you will use as your working directory inside Visual Studio Code: File → “Open Folder” → Select a folder

Inside the folder selected, add the structure below.

my-project/
├── CLAUDE.md
├── context/
│ └── content-context.md
├── profiles/
│ └── audience-profile.md
├──skills/
│ └── SKILL.md

Download the files here: Click on the green “Code“ button → download ZIP

Viewed inside VS Code, the structure will look like this:

This is just a basic Claude Code setup, though.

For more, check out our Claude Code series here.

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

When you switch from ChatGPT to Claude, the weird part isn’t the model.
It’s the reset.

You open a new chat, and everything you worked to build starts to disappear.

• The tone shifts.

• The format drifts.

• You end up re-explaining decisions you already made.

That’s not a prompt problem. It’s a memory problem.

Claude Code fixes this because you don’t “work inside a chat.”
You work inside a project, and the project loads your context from files every time.

This is the setup I wish I had from day one.

In this guide, you’ll get:

• 3 prompts to extract your existing chat history and turn it into reusable files

• A simple folder structure that keeps Claude Code consistent across every session

• A system you can set up in minutes and keep building on over time

No coding. You run the prompts in your current AI tool (ChatGPT, Gemini, etc), copy the outputs, and paste them into your Claude Code project.

Our series of guides on Claude, Claude Cowork and Claude Code keeps growing every week!

You can find all my Claude guides here. Consider becoming a paid subscriber if you like these guides 👇

Subscribe now

The shift that matters: from chat to project

Most people use AI like this:

Open a chat → ask for something → correct the output → repeat tomorrow from scratch

That works fine for one-off tasks.

However, once you start producing the same kind of work on a regular basis (guides, decisions, documents, analysis), two things start breaking:

• Inconsistency (each day it answers differently)

• Repetition (you keep re-explaining who you are, what you want, and how you work)

Claude Code fixes this because it reads from files you set up, not just your latest message.

So the goal isn’t to move your prompts over. It’s to move your structure.

It’s to build a stable structure Claude Code can load automatically.

Here is the workflow :

Instead of relying on the chat to "remember," you make the project carry the memory.

The minimum structure that actually scales

This setup is enough to make the system work — and simple enough to actually maintain. As your project grows, it grows with you.

my-project/
├── CLAUDE.md
├── rules/
│ ├── quality.md
│ └── boundaries.md
└── context/
├── you.md
├── project.md
└── writing-voice.md

Why does it work?

• It's not about having a lot of files. Each file just has one clear job.

• CLAUDE.md acts as the project’s memory.

• One thing worth knowing: CLAUDE.md doesn't replace your system prompt. It just adds context on top — things that are always true about your project. That's what makes it so consistent.

Think of it this way:

• CLAUDE.md = the essentials that always apply

• rules/ = longer policies and checklists (quality, boundaries, style)

• context/ = project facts, how you work, your voice, your audience

• skills/ = repeatable procedures (only when there is truly repetition)

This keeps you from stuffing everything into one file until it becomes a mess you can't manage.

Subscribe now

Prompts to extract your chat history and turn it into reusable files

The whole point is to pull out what you’ve already built across your history and turn it into project memory.

Here are three prompts, each focused on a different layer:

• Prompt 1: Extractor from a ChatGPT export

• Prompt 2: Project blueprint

• Prompt 3: Writing voice compiler (articles, emails, notes, and more)

The idea is simple: don’t try to do magic, do setup. You don’t need to run all three every time. Treat them like modules — use only what you need.

You run these prompts in your current AI chat tool.

What you get back are Markdown blocks you copy and paste into .md files in your Claude Code project.

Prompt 1: Extractor from a ChatGPT export

Read more

Hi!

I wrote today’s post in collaboration with . Ilia writes Prosper, a newsletter focused on practical AI workflows, systems, and frameworks for professionals who want to use AI more thoughtfully.

If you’re interested in building systems for working with AI, I highly recommend checking out Prosper.

Picture this: you need help with your business, so you open Claude (or any other AI) and start the same way: “Hey, quick context: here is what my business does, this is what I need help with today … ”

Claude helps you out. But then the next time comes, and you’re doing it all over again: rewriting context, re-explaining your situation, updating what was discussed last time.

Every. Single. Time.

That used to take up a lot of my time until I started using .md files. Once I did, I stopped re-explaining myself from scratch.

An .md file is a simple text file that holds context about you, your work, or your projects. You create it once. You reuse it everywhere.

In this guide, you’ll learn:

• How to create your own .md file

• 5 interview-style prompts that generate ready-to-use .md files for business context, writing voice, project work, content creation, and analysis

• How to use your .md file in Cowork

You can upload .md files to Claude, ChatGPT, or any other AI, but I recommend using it on Claude Cowork. In the next section, we’ll see why.

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

Cowork makes .md files more powerful

In regular Claude chat, you paste a .md file into the conversation. Works fine, but you do it every time.

Cowork is different.

Drop a .md file into a folder and Claude reads it automatically whenever you work in that folder. Open the folder, Claude already knows the context. No pasting. No re-uploading.

Here’s what that looks like in practice.

Ilia writes a newsletter called Prosper. His writing folder contains a voice profile with banned words, sentence patterns he avoids, positions he’d never take.

“When I open that folder in Cowork and ask for a draft, it doesn’t ask who I am. It reads the file and produces something that sounds like I wrote it on a Tuesday morning”

- Ilia Karelin

Same principle, any folder. Your client folder knows your delivery standards. Your research folder knows how you think about data.

Zero re-explaining. Claude starts already knowing your work.

Here’s how to use Cowork:

1. Download Claude for desktop

2. Open the app and choose “Cowork“

3. Give access to the folder containing the .md file

In the next section, we’ll learn how to create an .md file. Once you connect the file to Cowork, it’ll have it as a reference for future responses.

📚 Claude Cowork - Complete Guide: artificialcorner.com/p/cowork

📚 Move from ChatGPT to Claude: artificialcorner.com/p/switch-to-claude

📚 Best Claude Setup: artificialcorner.com/p/claude-setup

Subscribe now

How to create your .md file: the interview method

You don’t sit down and write the file completely from scratch. Claude interviews you and builds it from your answers.

Instead of staring at a blank document trying to articulate how you work, you let Claude interview you, and we’re going to give it a starting point.

It asks the questions. You answer honestly. At the end, it hands you a finished file.

The insight that makes this work: your profile is mostly about what you “reject”, not what you prefer. What you never want to see. What you’d never say.

That’s where the real signal is - not “I like direct communication” but “I’d never send a bullet-pointed email to a client.”

Here’s the base prompt:

You are going to help me create a context file I can use with Claude Cowork.

Ask me questions one at a time to understand my work: who I am, what I do,
how I like things done, what I never want to see in outputs, and what
“good work” means to me specifically. Use AskUserQuestion for it.

Ask me 40-50 questions total. One at a time. Push back if my answer is
vague — ask for a specific example. When we’re done, compile everything
into a .md file I can save and use with Claude.

Claude takes it from there.

Note: For the examples below, I told Claude to ask me 3 questions and then compile a first draft of the file. In practice, 40-50 questions produce a much richer profile — but this is a good way to see what you’re building toward before committing to the full interview. You can always run it again or ask Claude to update the file with more details later.

The key is answering specifically. “I want professional outputs” is useless. “I want outputs that sound like they came from a CFO - direct, numbers-first, no filler” gives Claude something it can actually apply.

In the previous prompt, we used AskUserQuestion to collect your preferences through preset options, but short multiple-choice answers only scratch the surface. Claude works best when it has rich, detailed context about you

For a more tailored experience, choose one of the five role-specific prompts below and answer as thoroughly as you can. The more detailed and thoughtful your responses, the more context Claude will have to work with.

📚 Claude Opus 4.6

📚 Claude - The Setup Most People Never Do

Prosper

Every week, practical AI workflows, prompts, and guides you can put to use the same day - no matter your technical background.

By Ilia Karelin

5 interview prompts that generate ready-to-use .md files

1. Business context

For owners, freelancers, and consultants who want Claude to help with anything business-related: client calls, contractor briefings, meeting prep, proposals.

You are going to help me create a business context .md file for Claude Cowork.

Ask me questions one at a time. Cover:

- What my business does - what I sell, who I serve, what problems I solve

- How I describe my business in my own words (not my website version)

- My team structure - employees, freelancers, partners, and who does what

- My biggest recurring challenges: meetings I dread, tasks I repeat, decisions I struggle with

- My communication style - how I write emails, how I run meetings, what I sound like

- My financial setup - how I track money, who I report to, what I need to know for accounting

- My non-negotiables - things I never want Claude to suggest or assume about my work

- What a great output actually looks like for me, with a specific example

Ask 40-50 questions. One at a time.

Interview rules:
- Push back on vague answers. If I say something generic, ask for a specific example.
- Call out contradictions. If something conflicts with an earlier answer, flag it.
- Don’t accept “I don’t know.” Reframe the question or approach it differently.
- Go deeper on interesting threads. If something unusual comes up, follow it before moving on.
- Ask for real samples - actual emails, documents, or decisions. Descriptions mean less than examples.
- Don’t validate my answers. No “great point” or “that’s interesting.” Just ask the next question or push deeper.
- If my answer sounds like how I wish I worked rather than how I actually work, call it out. Ask me for proof.
- If my answer could describe anyone in my profession, flag it. Ask what’s specific to me.

When we’re done, compile everything into a .md file with clear section headers.
Include the specific examples alongside the rules — they’re what make the file useful.

With this one, you will definitely see your prep time that used to take 20 minutes of re-explaining, now takes 2.

2. Writer’s voice

For newsletter writers, bloggers, and anyone who creates written content and wants Claude to draft in their actual voice — not a polished AI version of it.

You are going to help me capture my writing voice in a .md file.

Ask me questions one at a time. Cover:

- What I write and who I write for

- What my writing sounds like at its best — ask me to share an example paragraph

- What my writing sounds like when I’m just going through the motions

- Phrases I use naturally that feel like “me”

- Words and phrases I hate seeing in my own writing

- Topics I always come back to, even when I’m not trying to

- What I refuse to write about, or positions I’d never take

- My formatting instincts: long or short paragraphs? headers or prose? lists or sentences?

- Writers I admire — and specifically what I’d steal from each one

- What feedback I keep getting on my writing

- What feedback I keep ignoring, and why

Ask 40-50 questions. One at a time.

Interview rules:
- Push back on vague answers. If I say “I like directness,” ask me what that looks like in a specific paragraph.
- Call out contradictions. If something conflicts with an earlier answer, flag it.
- Don’t accept “I don’t know.” Reframe the question or approach it differently.
- Go deeper on interesting threads. If something unusual comes up, follow it before moving on.
- Ask for real samples — actual paragraphs I’ve written, sentences I’d never publish. Descriptions of style mean less than examples of it.
- Don’t validate my answers. No “great point” or “that’s interesting.” Just ask the next question or push deeper.
- If my answer sounds like the writer I wish I were rather than the writer I actually am, call it out. Ask me for proof.
- If my answer could describe any writer in my space, flag it. Ask what’s specific to me.

When we’re done, compile everything into a .md file with clear section headers.
Include the specific examples alongside the rules — they’re what make the file useful.

That’s the one I have in the folder I work in. Adjectives like “direct, warm, conversational” mean nothing to Claude. You end up with actual rules - hard nos, specific examples, the exact phrases you’d never use - the kind of guardrails that make Claude’s drafts sound like yours.

The rule that makes this work: 80% of a good voice profile is about what you reject. What sounds wrong. What you’d never publish. That’s what separates your voice from everyone else’s.

3. Client and project work

For consultants, agencies, and project managers whose Cowork folders hold client deliverables, project notes, and anything where working-style context matters.

You are going to help me create a client work .md file for Claude Cowork.

Ask me questions one at a time. Cover:

- The type of work I do with clients — deliverables, services, projects

- How I onboard clients and what I need to know before I start

- My delivery standards — what “done” looks like, how I format deliverables,

 what I never send without reviewing myself

- How I communicate with clients — tone, frequency, what I always and never say

- My biggest recurring frustrations with client work

- What makes a project go well vs. badly — in specific terms, not general ones

- How I handle scope creep, missed deadlines, or difficult conversations

- Tools and systems I use (where projects live, how I track things)

- What I’d want a smart assistant to know before helping me with any client task

Ask 40-50 questions. One at a time.

Interview rules:
- Push back on vague answers. If I give a process-level answer, ask me for a real example from a recent project.
- Call out contradictions. If something conflicts with an earlier answer, flag it.
- Don’t accept “I don’t know.” Reframe the question or approach it differently.
- Go deeper on interesting threads. If something unusual comes up, follow it before moving on.
- Ask for real samples — actual client emails, deliverable excerpts, or project decisions. Descriptions mean less than examples.
- Don’t validate my answers. No “great point” or “that’s interesting.” Just ask the next question or push deeper.
- If my answer sounds like how I wish I handled clients rather than how I actually do, call it out. Ask me for proof.
- If my answer could describe any consultant or PM, flag it. Ask what’s specific to me.

When we’re done, compile everything into a .md file with clear section headers.
Include the specific examples alongside the rules — they’re what make the file useful.

Drop this file into a client folder and Claude stops being a blank-slate assistant.

It knows your delivery standards, how you talk to clients, and what you’d never send without reviewing. One file, reused across every project.

4. Content creator

For YouTubers, podcasters, and social media creators who want Claude to help with ideation, scripting, and repurposing - in a way that actually fits their content.

You are going to help me capture my content identity in a .md file.

Ask me questions one at a time. Cover:

- What kind of content I make and where I publish it

- Who my audience is — not who I think they are, but what they actually respond to

- My content pillars: the 3-5 topics I return to again and again

- My on-camera or on-mic personality vs. my off-camera personality

- Content I’m most proud of and why it worked

- Content I’ve made that flopped — and what I think went wrong

- Content trends I refuse to follow even if they’d probably perform

- My production constraints: time, team size, tools I have access to

- The 3 creators I admire most and specifically what I’d take from each one

- What I want people to feel or do after consuming my content

- What I’d never make, even if someone paid me to

Ask 40-50 questions. One at a time.

Interview rules:
- Push back on vague answers. If I say “I want to be authentic,” ask me what that looks like in a specific piece of content I’ve made.
- Call out contradictions. If something conflicts with an earlier answer, flag it.
- Don’t accept “I don’t know.” Reframe the question or approach it differently.
- Go deeper on interesting threads. If something unusual comes up, follow it before moving on.
- Ask for real samples — actual video scripts, post captions, or content decisions. Descriptions mean less than examples.
- Don’t validate my answers. No “great point” or “that’s interesting.” Just ask the next question or push deeper.
- If my answer sounds like the creator I wish I were rather than the creator I actually am, call it out. Ask me for proof.
- If my answer could describe any creator in my niche, flag it. Ask what’s specific to me.

When we’re done, compile everything into a .md file with clear section headers.
Include the specific examples alongside the rules — they’re what make the file useful.

Claude stops suggesting ideas that don’t fit your channel. Scripts stop sounding like they could belong to anyone.

When you repurpose a video into a thread or a newsletter, the thing that made the original work doesn’t get lost in translation.

5. Analyst and researcher

For data analysts, researchers, and strategists whose Cowork folders contain reports, datasets, and research notes where output format and analytical standards matter.

You are going to help me create an analyst context .md file for Claude Cowork.

Ask me questions one at a time. Cover:

- What I analyze and for whom — who reads my outputs and what they do with them

- How I define a good analysis — what questions it answers, what format it takes

- My output preferences: how I structure reports, what I always include, what I always cut

- The frameworks or mental models I use most, even informal ones

- What “rigorous” means to me — where I draw the line on confidence and evidence

- My biggest recurring frustrations with how analysis gets done or presented

- Tools and data sources I work with regularly

- What I want someone to do after reading my analysis: decide, act, question, understand?

- What I never want to see in an analysis output — vague language, unsupported claims,
specific phrases that signal lazy thinking

Ask 40-50 questions. One at a time.

Interview rules:
- Push back on vague answers. If I describe a preference, ask me for an example of when it mattered in real work.
- Call out contradictions. If something conflicts with an earlier answer, flag it.
- Don’t accept “I don’t know.” Reframe the question or approach it differently.
- Go deeper on interesting threads. If something unusual comes up, follow it before moving on.
- Ask for real samples — actual report excerpts, analysis decisions, or data interpretations. Descriptions mean less than examples.
- Don’t validate my answers. No “great point” or “that’s interesting.” Just ask the next question or push deeper.
- If my answer sounds like how I wish I analyzed rather than how I actually do, call it out. Ask me for proof.
- If my answer could describe any analyst in my field, flag it. Ask what’s specific to me.

When we’re done, compile everything into a .md file with clear section headers.
Include the specific examples alongside the rules — they’re what make the file useful.

No more “here are the findings” summaries that don’t match how you think or how your audience reads.

Outputs formatted and reasoned the way you’d do it yourself - because the file tells Claude exactly what that looks like.

Subscribe now

Do This Now

Follow these steps:

1. Open the Claude app

2. Go to the Cowork tab.

3. Paste one prompt from the list above into Cowork

4. Answer every question.

When Claude pushes back and asks for specifics, give them. That’s where the file gets good. When it hands you the finished file, save it with a descriptive name - [voice_profile.md], [business_context.md], whatever fits what you built.

Drop the .md file in your most-used work folder and give Cowork access to that folder:

At the end of your first prompt, type this:

[your prompt]

Before you start, read [your filename].md first

Here’s an example from just a couple of words I wrote in Cowork:

Done.

Every time you open that folder and give Cowork a task - prep a meeting, draft an email, analyze a dataset - it reads the context file before doing anything. It already knows who you are.

If you want to take this to the next level, use the scheduled tasks feature:

📚 Claude scheduled tasks: artificialcorner.com/p/claude-automates

Subscribe now

Don’t forget to subscribe to Prosper, where you can find practical AI workflows, systems, and frameworks for professionals who want to use AI more thoughtfully.

Here are some of Ilia’s best posts:

• Commands vs Skills vs Agents in Claude Code

• Official Claude Prompting: Everything You Need To Know

• I Stopped Scrolling for AI News

I used to spend hours writing scripts to automate the web, Excel reports, etc.

Thanks to Claude, now anyone can do this in minutes without writing code!

The best part? You can schedule tasks directly in Claude, so it runs them automatically on a recurring basis. Set it once and it just works, while you focus on everything else.

In this guide, I’ll show you step by step how to use Claude to schedule tasks like:

• Weekly research: Track topics, competitors, or industry news

• Daily reports: Compile data from spreadsheets into a formatted summary

• Daily file organizer: Sort and clean up files in your downloads folder

For this, we’ll use two tools: Claude in Chrome to automate the web, and Claude Cowork to automate your files.

Twice a week, I share the best ways to get more out of Claude.

Click here to read our best Claude guides

and consider becoming a paid subscriber if you like my Claude guides 👇

Subscribe now

Claude in Chrome: Automate the Web

Claude in Chrome is a Chrome extension that automates the web. To work with it, follow these steps:

1. Open Google Chrome

2. Go to Claude in Chrome → Click “Add to Chrome” to install

3. Sign in with your Claude account

4. Grant permissions so Claude can interact with your browser

The Claude icon will appear in your Chrome toolbar. Click it to open Claude in a side panel that stays visible while you browse.

Anything that involves repetitive web actions can be automated with this extension.

Just click the Claude sidebar while browsing any site, describe the task, and Claude will control your browser until it’s done.

Here’s one way Claude in Chrome has automated my work.

Every week, I track which Substack newsletters break into the top 200. I analyze what's working and spot potential collaborations.

To get data for my analysis, I have Claude scrape Reletter’s Substack charts with this prompt:

Build a table with the following columns for the first 250 elements in the table:

• Ranking

• Newsletter

• Date (today’s date)

• Type = free/paid

Give me a link to download the file in csv format

Here’s a video demo.

If you do this manually (clicks, copy-paste, etc), it takes forever. With Claude, it takes a couple of minutes.

It gets even better with the scheduled tasks feature. You can turn a prompt that works into a task that runs any time you want. Just click the three dots (at the top of the chat) and select “Convert to task.“

You’ll be redirected to the Claude in Chrome settings with the window below. Mine is empty, but if you clicked on “convert to task,“ the fields should be automatically filled based on your prompt.

Claude might tweak your prompt to make it more robust. Review it and adjust if necessary. Also, toggle on “schedule” and pick when the task should run.

Once you’re happy with it, click on “Create shortcut.“

To edit a shortcut later: Go to Claude in Chrome settings (under the three dots) → Settings → Shortcut

You’ll see all your scheduled tasks listed there. I schedule mine to run every Tuesday.

Now every week, Claude scrapes the site and downloads the data for me.

That’s how you automate the web. Now let’s see how to automate your desktop.

Important note: Scheduled tasks only run while your computer is awake

Subscribe now

Cowork: Automate Your Files

Claude Cowork feels like a virtual employee, capable of managing files, organizing data, and performing multi-step tasks directly on your computer.

Recently, Cowork also got scheduled tasks!

It works a lot like Claude in Chrome. Write a prompt once, and Cowork can turn it into a scheduled task.

Here’s how to create a scheduled task from scratch:

1. Download Claude for desktop

2. Open the app and choose “Cowork“

3. Click “Scheduled” in the left sidebar

4. Click “+ New task”

5. Fill in the details (prompt, task name, etc)

6. Click “Save” to add a new task to the Scheduled tasks page.

We can also turn a working prompt into a task that runs whenever we want.

In the previous example, I used Claude in Chrome to set up a task that runs every Tuesday. It scrapes the data and generates a fresh dataset with a Substack newsletter ranking.

Now I’ll have Cowork analyze that data and generate a weekly report showing the 10 newsletters with the biggest gains and biggest drops.

Here’s the prompt I use:

Based on the data collected in the folder “Substack Newletter Ranking” generate an analysis that covers:

• Biggest rank gains: Top 10 newsletters with the largest improvement (largest drop in rank number)

• New entrants: Top 20 newsletters that appear in the rankings only in the most recent dataset

• Biggest rank losses: Top 10 newsletters with the largest decline (greatest increase in rank number).

The folder contains two types of datasets:

• Paid: rankings based on ARR

• Free: rankings based on number of free email subscribers

Create the same analysis separately for each dataset (Paid and Free)

After running the task, Cowork drops a spreadsheet right in my folder with the full analysis. Congrats to the newsletters climbing fast … hope to be on the list soon :)

Now I can turn this working prompt into a scheduled task.

In the chat, click the dropdown arrow and select “schedule.”

The /schedule command will appear. Type something like “run this every [day] at [time]“

My previous automation generates datasets on Tuesday around 5 PM, so for this one, I’ll schedule it a bit later: “run this every Tuesday at 5:10 PM.”

Cowork will ask for confirmation. Just press enter.

Your scheduled task has been created!

To view and manage all your scheduled tasks, open the left sidebar and select “Scheduled.”

Important note: Scheduled tasks only run while your computer is awake and the Claude Desktop app is open.

Congrats! Now you know how to schedule tasks in Claude.

Check out the guide below for more Cowork use cases.

Final thoughts

• Not everything should run on autopilot. Review your chat history and identify which tasks actually benefit from being scheduled

• You can schedule as many tasks as you like (Claude will happily work in the background), but each one draws from your daily usage allowance. The $20 Pro plan may be enough if you’re running a few lightweight tasks, but heavier workflows might call for the $100 or $200 tier

• If this was helpful, consider becoming a paid subscriber. It helps me move up the Substack rankings (like my fellow writers!) and keeps these guides coming

I’m thinking of making a guide packed with Claude and Cowork use cases. If there’s a specific one you’d like covered, drop it in the comments.

Click here to read the best Claude guides

Subscribe now

You know Claude is best at following instructions.
You want to try Claude’s skills and Cowork

But you have too much built up in ChatGPT to just walk away. You don’t want to lose all your chats, memory, custom instructions, etc, so it feels like you’re trapped.

In this guide, we’ll fix that.

In a few minutes, you’ll learn the best ways to move from ChatGPT to Claude:

• Option #1: Transfer ChatGPT memory to Claude (with Anthropic’s prompt)

• Option #2: Transfer ChatGPT projects, GPTs & custom instructions to Claude

• Option #3: Transfer all your data from ChatGPT to Claude

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

Option #1: Transfer memory to Claude (Anthropic’s prompt)

Anthropic just released a method to transfer your preferences and context from other AI providers to Claude.

You only have to follow two steps.

Step 1: Copy Anthropic’s prompt below and paste it into ChatGPT

I’m moving to another service and need to export my data. List every memory you have stored about me, as well as any context you’ve learned about me from past conversations. Output everything in a single code block so I can easily copy it. Format each entry as: [date saved, if available] - memory content. Make sure to cover all of the following — preserve my words verbatim where possible: Instructions I’ve given you about how to respond (tone, format, style, ‘always do X’, ‘never do Y’). Personal details: name, location, job, family, interests. Projects, goals, and recurring topics. Tools, languages, and frameworks I use. Preferences and corrections I’ve made to your behavior. Any other stored context not covered above. Do not summarize, group, or omit any entries. After the code block, confirm whether that is the complete set or if any remain.

After a few seconds, you’ll get the information ChatGPT has been collecting about you in its memory across every chat you’ve had: who you are, what you do, what you like/dislike, and other preferences.

Step 2: Copy and paste the results into Claude’s memory settings.

Done! Your ChatGPT memory has been transferred to Claude

But that’s only the memory!

If you have other important information in your custom instructions, projects, GPTs, and chats, options 2 and 3 will come in handy.

Subscribe now

Option #2: Transfer GPTs, Projects & custom instructions

This is my favorite way to move from ChatGPT to Claude.

Why? Because most of us don’t need every single thing we’ve shared with ChatGPT. If you go with option #3, you’ll get a massive .zip file with all your ChatGPT data. There are workarounds to connect this file to Claude, but they come with some downsides:

• ChatGPT might’ve saved inaccurate info about you in its memory

• Both useful and useless chats will get exported (which might confuse Claude later)

For most users, transferring the following will be enough:

GPTs & Projects → Claude Projects

If you’ve created your own GPT or project and want to transfer it to Claude, there’s an easy way to do it: transfer its instructions to a Claude project

1. Go to your GPT or Project

2. Copy the GPT/Project instructions

3. Download the GPT/Project knowledge files

4. Go to Claude → Projects (left bar) → New Project

5. Paste the instructions and upload the files

If you want to transfer your project’s memory, do this:

1. Go to your ChatGPT project

2. Use this prompt “Based on everything discussed in this project, create a comprehensive summary and context“

3. Copy the response

4. Go to your Claude project

5. Click on the + inside Files → Add text content → Paste the response → Click “Add content”

ChatGPT Memory / Custom Instructions → Claude Memory

You can quickly transfer ChatGPT’s memory and custom instructions to Claude. Here’s how.

1. Click on ChatGPT settings

2. Go to Personalization

3. Scroll down to Memory → Manage (or Custom Instructions → copy text)

4. Copy your saved memories (tip: read the memory and verify that the info about you is accurate. If something doesn’t sound quite right, delete it)

5. Go to Claude settings → Capabilities → Turn on Memory

6. Open a new chat → Type “Update your memory about me with this“ → Paste the saved memories (or custom instructions)

That’s it!

Now Claude has the full memory of what ChatGPT has learned about you over the years. If you want to review or edit your memory in Claude, go to Settings → Capabilities → Memory → Memory from your chats

📚 If you’re new to Claude, you might find these guides useful:

• Best Claude features: artificialcorner.com/p/claude

• Don’t use Claude wrong: artificialcorner.com/p/claude-levels

• How Claude Cowork saves me time: artificialcorner.com/p/cowork

Option #3: Transfer all your data from ChatGPT to Claude

This method involves exporting your ChatGPT data and uploading it to Claude.

1. Go to ChatGPT settings

2. Select “Data controls”

3. Click on export data

After some hours, you’ll get an email with your ChatGPT data (I got mine after 24 hours)

You’ll get a zip file. Download it and unzip it.

The data export from OpenAI is big (mine was 3.5GB). Here are the two best ways to connect it to Claude:

• Create a Claude project and upload only the chat.html file

• Point Claude Cowork to the whole folder

#1 Upload only the chat.html file to a project

chat.html is a smaller file inside the data export that contains our chat history

We can upload this file to a Claude project. The only issue is that project files have a 30MB limit in Claude.

If your chat.html file is under 30MB, do this:

1. Go to Claude → Projects (left bar) → New Project

2. Click on the + inside Files and upload the chat.html file

If your chat.html file is over 30MB, do this:

1. Open the chat.html file

2. Select all the text (Ctrl + A) and copy everything

3. Go to Claude → Projects (left bar) → New Project

4. Click on the + inside File → Add text content → Paste the text → Click “Add content”

After following the steps, Claude will have knowledge of your ChatGPT chats.

#2 Point Claude Cowork to the whole folder

Cowork is a tool available in the desktop app. It’s different from regular Claude.

One key difference is that Cowork can access your local files directly (without manual file uploads).

Although Claude Cowork can’t connect to your Claude memory or projects, it’s still useful when you want to find specific information in your ChatGPT data.

Just download Claude for desktop, install it, and open the app.

Choose Cowork and give it access to the folder that contains your ChatGPT data.

After that, Cowork will have all your ChatGPT knowledge.

That said, regular Claude still doesn’t have this knowledge. You could ask Cowork to create a blueprint from your ChatGPT data and then upload it to Claude's memory. If you need that, here’s a Notion repository where you can find a long migration prompt that generates that blueprint.

Subscribe to get my FREE Claude course.

20+ lessons delivered straight to your inbox 👇

Subscribe now

The people getting the most out of Claude Code aren’t better at prompting.

They build systems.

If you use AI today the way most people do, you probably follow the same loop:

Open a chat → write a prompt → fix the output.

Then the next day, you start from scratch.

You’re not building an assistant. You’re improvising. Over time, that has a real cost:

• Results become inconsistent

• The tone shifts from one request to the next

• You keep re-explaining things you thought were already settled

Claude Code is different.

Instead of a chatbot that just replies, it works as an environment that can read files, run tools, generate documents, and improve outputs — all inside a single project.

That shift from chat to project is what stops everything from restarting at zero every time you open it. This is also why building a CLAUDE.md file matters (think of it as the project’s memory)

You design it once, and it shapes every interaction with Claude from that point on. You get consistent judgment, the same response structure, and clear boundaries.

By the end of this guide, you'll have:

• A practical definition of what CLAUDE.md is and what it actually changes

• A simple checklist for what to write so Claude stops improvising

• CLAUDE.md templates you can copy-paste (and adapt)

• A plug-and-play, interview-style prompt that generates your own CLAUDE.md

Our Claude Code series grows every week with a new guide.

Click here to read our best Claude guides

And consider becoming a paid subscriber if you like my Claude guides 👇

Subscribe now

What is CLAUDE.md and why it matters

A CLAUDE.md file is, quite literally, a Markdown file.

Plain text. Simple structure. Designed for writing instructions that are readable and easy to organize.

It is not code. You don’t need to know how to code to use it.

In Claude Code, CLAUDE.md acts as the project’s memory. It’s a special file that Claude reads at the start of every conversation to:

• Load your rules

• Follow established conventions

• Keep context consistent across sessions

The goal isn’t documentation. The goal is to operate.

Claude gets a stable framework — so it stops improvising every time you open a new chat.

Think about the difference between these two scenarios:

• A chat with no memory: each session depends on what you remember to ask for, and how clearly you phrase it that day

• A project with memory: your “way of working” lives in files, not in your head

In Claude Code, that memory isn’t just one file. It’s a hierarchy, and it loads automatically when you start. That hierarchy also determines which rules carry more weight.

One thing worth knowing: CLAUDE.md doesn’t replace Claude Code’s default system prompt. It gets added as consistent content inside the conversation — on top of the base prompt, not instead of it.

That might sound like a small detail. It’s not.

It means CLAUDE.md is one of the most stable levers you have because:

• It’s present in every session of your project

• It’s written as rules your whole team can read and scale

Subscribe now

What to put in your CLAUDE.md so Claude stops improvising

My recommendation is to keep it:

• As human as possible

• Compact, so it can scale over time

• Focused

This content loads in every session. Every word counts.

Here’s the simple version:

Identity: who you are here and what you are here to do

Defining the role is not a small detail. It’s how you control behavior.

Your identity section should answer three things:

• Role: editor, PMO, manager, writer

• Goal: the outcome you want consistently

• Operating principles: how you prioritize and how you handle uncertainty

The more specific you are here, the less Claude improvises.

Output: a delivery contract that stays stable

Instead of asking “write me a report” every time, define the structure once in CLAUDE.md. For example:

• Summary in at most X lines

• Risks and constraints

• Compared options

• Next steps

In practice, this cuts down revisions. Claude stops picking a new format every time it responds.

Quality: how the work is checked

Claude performs much better when it knows what “correct” means. Give it criteria, expected outputs, or examples. Lines like:

• Include an introduction that previews the value

• Use concrete examples

• Do not invent data — ask for it instead

• Avoid claims you cannot verify

Boundaries: what Claude should not do, even if it sounds reasonable

Claude Code already has a permission system. But your CLAUDE.md should set its own boundaries too — it reduces risk and prevents drift.

Some non-technical boundaries that work really well:

• Do not assume numbers, dates, or policies if they are not in the files

• If critical information is missing, ask before proposing a final plan

• Separate facts from hypotheses

• Avoid recommendations with material impact unless you clearly present risks and trade-offs

Navigation: where the truth lives inside the project

This one surprises people who don’t code.

AI does not guess your context. It reads it.

Claude Code lets you pull in files from CLAUDE.md using @route syntax. This keeps CLAUDE.md from turning into an endless file — and tells Claude exactly where to look.

Instead of pasting three pages of brand guidance in CLAUDE.md, keep a separate file and reference it:

• Brand guide: @docs/brand-voice.md

• Glossary: @docs/glossary.md

• Article template: @docs/article-template.md

This keeps the whole design modular (much easier to maintain long-term)

Example structure:

my-skills/
├─ CLAUDE.md
├─ rules/
│  ├─ rules.md
│  └─ boundaries.md
└─ skills/
   └─ weekly-decision-brief/
      └─ SKILL.md

Here’s a practical rule to keep CLAUDE.md minimal, while still making it scalable:

a) Identity

• Keep in CLAUDE.md: the “who you are here” that should always be present

• Move to context/: long project history, quarterly goals, detailed audience definitions

b) Output

• Keep in CLAUDE.md: the default format you expect most of the time

• Move to skills/: repeatable procedures, or anything you want to trigger by intent instead of running on every request

c) Quality

• Keep in CLAUDE.md: 3 to 7 universal rules that apply across the project

• Move to rules/quality.md: longer checklists, anything you want to version and maintain in parts

d) Boundaries

• Keep in CLAUDE.md: simple, universal project boundaries (usually 2 to 6 bullets)

• Move to rules/boundaries.md: specific policies, or when you want boundaries isolated so they don’t get diluted

e) Navigation

• Keep in CLAUDE.md: high-level pointers about where things live

• Move to context/: glossaries, sources of truth, templates, examples, assets

How to keep CLAUDE.md current without letting it degrade

If you write CLAUDE.md once and never touch it again, you will miss most of its value.

A better approach:

• Review it on a regular basis

• Keep “memory” specific, structured, and easy to maintain

A simple and realistic cycle looks like this:

1. Start with /init to generate an initial CLAUDE.md based on your project. Then edit it so it reflects what the project actually needs.

2. When you notice the same mistake happening again, don’t fix it only inside a one-off prompt. Turn it into a rule or import a reference file Claude can reuse.

3. If the file grows, move rules into a rules/ folder and split them by topic.

4. For anything that doesn’t apply every time, use skills. Skills load only when needed, so they don’t consume context in every session (I rely on this a lot)

One more thing: keep security in mind. Design your CLAUDE.md with clear boundaries, and avoid placing sensitive information inside it.

Subscribe now

CLAUDE.md templates you can copy-paste (and adapt)

CLAUDE.md should be short, readable, and practical. Anything too specific or too heavy belongs in imported files or separate rule files.

Here are three templates you can use:

📄 CLAUDE.md template for decisions and briefs

my-skills/
├─ CLAUDE.md
├─ rules/
│ ├─ rules.md
│ └─ boundaries.md
└─ skills/
 └─ weekly-decision-brief/
 └─ SKILL.md

# Identity
You are my thinking partner for executive decisions.
Prioritize clarity, trade-offs, and risk. Do not fill space with empty opinions.

# How to work
- If critical information is missing to make a decision, ask questions first.
- Separate facts (given) from hypotheses (inferences).
- Do not invent numbers, dates, or policies.

# Output standard: Decision brief
ALWAYS deliver using this structure:
1) Context (5 to 7 lines)
2) Key signals (max 7)
3) Risks and constraints (max 7)
4) Options (min 2, max 4) with pros and cons
5) Open questions (always)
6) Next steps (concrete actions)

# Navigation
- Reasoning rules live in: rules/rules.md
- Boundaries live in: rules/boundaries.md
- Decision outputs live in: skills/weekly-decision-brief/SKILL.md

Why it works:

• It uses specific rules, a clear structure, and imports for heavier context — no bloat in the main file

• CLAUDE.md is read at the start of every conversation, so the restart cost drops to almost zero

📄 CLAUDE.md template for writing articles in your voice

claude-md-article/
├─ CLAUDE.md
├─ context/
│  ├─ audience.md
│  └─ article-goal.md
├─ rules/
│  ├─ output.md
│  ├─ quality.md
│  └─ boundaries.md
└─ skills/
   └─ article-outline/
      └─ SKILL.md

# Identity
You are my editor and content strategist.
Your job is not only to write, but also to shape an article that is useful and actionable.

# Editorial rules
- Avoid generic lines. Use concrete, verifiable examples.
- If something is not supported by a source or a project file, say so explicitly.
- Keep the logic clear: problem, framework, example, application.

# Required structure
- Introduction: sell the article (the promise plus what the reader will achieve).
- Maximum of 4 main sections (not counting the intro).
- Must include at least:
- 1 diagram
- 1 mind map
- 2 applied examples based on real scenarios

# Navigation
- Audience: context/audience.md
- Goal and angle: context/article-goal.md
- Output contract: rules/output.md
- Quality checks: rules/quality.md
- Boundaries: rules/boundaries.md
- Article Outline Skill: skills/article-outline/SKILL.md

rules/output.md:

# Output Contract
- Strong intro: the promise plus what the reader will get
- Max 4 main sections (excluding intro)
- Must include:
- 1 diagram
- 1 mind map
- 2 concrete examples (include "Why it works")

rules/quality.md

# Quality Checklist
- No generic filler
- Every concept has a concrete example
- Separate facts from assumptions
- If something is missing, ask or state the gap

Why it works:

• CLAUDE.md stays lightweight — identity and pointers only, nothing competing with the actual content

• Output and Quality live in rules/, so they’re easy to version and update independently

• The Skill is called only when needed (outline and structure), so it doesn’t eat up context on every request

📄 CLAUDE.md template that turns meeting notes into a plan

Good starting point if the project is new and there isn't much complexity yet.

# Identity

You are my operations assistant.

Your priority is to turn ambiguous information into clear actions.

# Information discipline

- Separate confirmed decisions from open items and risks.

- If an owner or a date is missing, ask for it. Do not invent it.

# Required output: Action plan

1) Meeting summary (max 10 lines)

2) Confirmed decisions

3) Actions (mental table: what / who / when / definition of done)

4) Dependencies and blockers

5) Questions to close gaps

# Boundaries

- Do not assume missing data

- Avoid speculation

# Navigation

If this grows, split it into rules/ and skills/

Why it works:

• It reduces friction at the start. No over-engineering before you need it

• As the system grows, you can migrate cleanly into the same architecture (rules/, context/, skills/)

Generate your own CLAUDE.md with this prompt

There’s a faster and more effective way to create a CLAUDE.md: let Claude Code run an interview with you. At the end of the interview, it’ll generate a CLAUDE.md and every file your project needs.

By answering simple interview questions, you’ll define:

• who Claude is in your project

• how it should respond

• the quality bar it must meet

• the boundaries it must respect

• how it should navigate the rest of your files

Think of it as guided onboarding. You answer once, and those answers give Claude Code enough structure to build your full system.

After that, you never have to explain it again.

Using it is straightforward:

1. Open a Claude Code chat

2. Paste the interview prompt below

3. Answer one question at a time. Take your time (it’s worth it!)

When the interview is done, type:

Now generate my folder structure and all files

Claude returns each file’s content, ready to copy-paste into your project or generate directly in your workspace.

Paste this prompt into Claude Code (67 interview questions)

If you want more guides like this, support us by becoming a paid subscriber today and get full access to our best guides and other perks.

Subscribe now

You are Claude Code acting as an interviewer.

Goal: run a structured interview to generate my project system:

- CLAUDE.md

- rules/output.md

- rules/quality.md

- rules/boundaries.md

- context/project.md

- context/audience.md

- context/glossary.md (if needed)

- context/sources.md (if needed)

- 1–3 skills/*/SKILL.md based on repeatable workflows

- optional: examples/, templates/, assets/

Interview rules:

- Ask ONE question at a time.

- Wait for my answer before continuing.

- If my answer is vague, ask a short follow-up (max 1 follow-up).

- Accept “N/A” and “Not decided yet” without pushing.

- Do not write the files until the interview is complete.

After the interview:

1) Summarize what you learned in 8–12 bullets.

2) Propose a folder structure.

3) Output each file in a separate block labeled with its path, like:

[CLAUDE.md]

…content…

[rules/output.md]

…content…

Then start the interview with Question 1.

Here are the interview questions (do not show them all at once; ask one by one):

A) Identity and purpose

1) What’s the name of this project (one line)?

2) What are you trying to produce in the end?

3) What’s the main outcome you want, in one sentence?

4) What role should Claude play in this project?

5) What are the top 3 responsibilities Claude must handle every time?

6) What are 3 things you never want Claude to do in this project?

7) What should the default tone feel like? Pick 3 adjectives.

8) What tone or style should Claude avoid completely?

B) Quality bar

9) What does a “perfect” response look like here? Name 3 concrete traits.

10) What makes a response bad or unusable? List red flags.

11) What 5 criteria will you use to judge quality?

12) When tradeoffs happen, what should Claude prioritize? Put in order.

13) How rigorous do you want this by default (practical / medium / high)?

C) Output defaults

14) What output format do you want by default?

15) What sections must show up every time?

16) What sections should appear only sometimes?

17) What should the default length be, and when should it switch?

18) What types of examples should Claude include?

19) What kind of ending do you prefer (next steps / checklist / open questions / decisions)?

D) Uncertainty

20) If key info is missing, what should Claude do first?

21) What is Claude allowed to assume without asking?

22) What should Claude never assume?

23) How should Claude label uncertainty?

E) Boundaries

24) What topics are out of scope or off-limits?

25) Any confidentiality/compliance constraints?

26) What decisions should Claude not make for you?

27) Any style boundaries?

28) Any operational limits (tools/links/workflows allowed)?

F) Project context

29) Describe the project context in 8 bullets.

30) What parts of context change often?

31) What parts are stable truths?

G) Audience

32) Who is the end reader/user?

33) Biggest pain points?

34) What are they trying to do quickly?

35) What confuses them?

36) What examples would feel instantly relevant?

H) Vocabulary and sources

37) Mini glossary: 10 key terms + one-line meanings.

38) Trustworthy sources?

39) Sources to avoid?

40) What claims must be backed up or clearly qualified?

I) Navigation

41) Where should rules/context/templates/examples/assets live?

42) What’s the read-first order of files?

43) Which files are required vs optional?

J) Skills (repeat 1–3 times)

44) What’s one repeated task that drains time/energy?

45) Typical input?

46) Ideal output?

47) Common failure modes and prevention?

48) Trigger: when to use / when not to use?

49) 3 success criteria?

50) Want a second Skill? (repeat 44–49)

51) Want a third Skill? (repeat 44–49)

K) Examples/templates

52) Provide 2 real input samples.

53) Provide 2 ideal output samples.

54) What reusable template do you want?

55) Best before/after transformation?

L) Diagrams/mind maps

56) 2 ideas that need a diagram?

57) 2 ideas that need a mind map?

58) Preferred visual style?

59) What should never show up in visuals?

M) What stays in CLAUDE.md vs moves out

60) 5 rules that always apply, no exceptions?

61) Which rules are situational, and what do they depend on?

N) Defaults

62) Default format if user doesn’t specify?

63) Default depth if user doesn’t specify?

64) Default behavior if critical context is missing?

O) Source-of-truth hierarchy

65) If files disagree, what’s the hierarchy?

66) Which files are source-of-truth vs reference-only?

P) Skill triggers

67) What repeated work deserves a Skill, and how will you recognize it?

Start simple. With this interview, you can build a clean, functional CLAUDE.md in minutes.

Then, as your project grows, you can keep moving specific details into rules/, context/, and skills/ without breaking anything.