“Think step by step.”
“You MUST.”

Older models needed these. Claude doesn’t need them anymore, and some of them just slow it down.

Claude can now do things it couldn’t before. It works on a task for hours, builds a finished file you can send, and more.

Those need a different kind of prompt.

I read Anthropic’s docs on how to prompt Claude so you don’t have to.

Here’s what still works, what you can drop and what’s new, from beginner to pro.

In this guide, you’ll learn:

• The 6 parts of a good prompt, and the 2 old habits to drop

• How to prompt Claude for real work and long sessions

• The instructions pros write once and never repeat

• The anatomy of a Claude prompt, in one image you can save

Share

1. Beginner: the basic prompt formula

Start here even if you’ve used Claude for a while. The anatomy at the end of this guide starts with these six parts.

The 6 elements

1. Task context: tell Claude what it’s doing and, if it helps, who it’s playing. One sentence is enough. A role mostly sets the focus and the tone of the answer.

Weak: Rewrite my resume

Strong: You are a career coach with 15 years of experience. Help me improve my resume

2. Tone: the easiest way to set it is to say who the answer is for.

Weak: Explain quantum computing

Strong: Explain quantum computing to a high school student with no physics background

3. Rules, with the reason: when Claude knows why a rule exists, it follows it better, and it handles the cases your rule didn’t cover.

Weak: Keep it short.

Strong: Keep it under 150 words. I’ll read it on my phone between meetings.

4. Examples: still one of the most reliable ways to get the format and tone you want. Anthropic recommends 3 to 5, different enough that Claude doesn’t copy one of them.

Write [the thing you need]. Match the style of these examples:

[paste 3 to 5 examples]

5. Attach the image: if you have a chart, a screenshot or a slide, attach it. Don’t retype the numbers. Opus 5.5 reads images more accurately than the models before it, including things that depend on where they sit (which box an arrow points to, when a meeting starts on a calendar screenshot).

6. Ask for the finished file: say which file you want and what goes in it. Claude can make the spreadsheet or document itself, so you don’t have to copy its answer out of the chat and reformat it.

Make this a spreadsheet: one row per vendor, with columns for cost, contract end date and owner.

Put the six together and you get a prompt like this:

You’ll rarely need all of it. A quick question needs a single line. But when a task matters, go down the list and see what’s missing.

One habit to add

Let Claude ask you questions first. End any important prompt with the line below

Before you start, ask me any questions you have.

Claude then asks what it needs before it starts. When it built my contributors’ payout system, two of its questions landed on bugs that were already hiding in the old spreadsheet.

The 2 habits to drop

Older models needed these. Today’s models don’t, so you can stop writing them.

1. “Think step by step” and “take a deep breath”: Opus 5.5 thinks before every reply, and it decides how much on its own. Anthropic tested this in a chat product: removing a “think carefully” line made replies start sooner, with no clear drop in quality.

2. Shouting at Claude: In an old prompting guide, one of my principles was to write “You MUST” in your prompts.

Older models skipped instructions, so we shouted. Newer ones follow instructions closely, so you don’t need to anymore. Anthropic now tells people who build with Claude to dial it back: “Use this tool when...” instead of “CRITICAL: You MUST use this tool when...”

In a chat, a “MUST” won’t break anything. It just isn’t needed. Write your rules the way you’d say them to a colleague.

Your next step

Pick one thing you’ll ask Claude this week: an email, a summary, a plan. Before you send it, go down the six elements and add the ones that are missing. Then end with the questions line.

Subscribe now

2. Intermediate: real work and long sessions

The first tips below help on any task. The rest are for long tasks.

All of them work the same in Claude Cowork or in Claude Code.

Any task

1. Name the defaults you don’t want: with no direction, Claude falls back on the same few looks and phrases. A list of specific things to avoid works much better than asking for something “less generic.”

Example 1: Design

“Make it look less generic” mostly swaps one default look for another. Name what to leave out instead:

Create [the deck, page or chart you need].

Don’t use [a look you don’t want], [something you always delete] or [a layout habit, like more than three bullets per slide].

If you don’t like what it picked instead, add that to the list and ask again.

Example 2: Writing

Same idea. Add the words and habits that don’t sound like you:

Write [what you need].

Don’t use [words you’d never use], [phrases you hate] or [a habit you don’t want, like em dashes].

Most of a good voice file is about what you reject. “I’d never use semicolons because they make my writing sound like a college essay” tells Claude more than “I like direct writing.” Save that list in a .md file once, and you won’t have to type it in every prompt.

2. Ask it to mark what it couldn’t confirm: for research and analysis, “I couldn’t find this” is worth reading. Asking for it makes it easy to spot.

Add this to your prompt:

Mark anything you couldn’t confirm, and say where you looked.

Long tasks

Opus 5.5 keeps going on long, multi-step work much better than the models before it. That changes what a good prompt needs.

1. Say what “done” looks like: give Claude the whole task in one message, say what finished looks like, and tell it when you want it to stop and ask you. Then leave it alone until it’s done.

Here’s a template you can fill in:

[The whole task]

Done means: [what finished looks like].

Stop and ask me only if [the call you want to make].

2. Add to a running task: if you remember something halfway through, send it as a follow-up while Claude works. Runs are longer now, so starting over costs more.

Type it like any other message:

Also [what you forgot to ask for].

Claude picked it up without starting over. Its summary at the end said where the new slide went:

3. Keep the task list in a file: for a task that will take a while, ask Claude to keep its to-do list in a file in your folder. A long run fills the context window (Claude’s working memory for the chat), and Claude then summarizes older turns. A list in a file survives that, and it shows you at a glance what’s done and what’s left.

Add this line to your prompt:

Keep a checklist in tasks.md. Tick each item when it’s done, and add anything new you find.

4. In a project, say when answers are settled: in a long chat, Claude sometimes goes back over an earlier answer while it thinks about a short follow-up. That slows the reply. If your follow-ups feel slow, add this to the project’s instructions:

Once you have answered something, treat that answer as done. Focus on what I’m asking now, and don’t go back over an earlier answer unless I ask about it or point out a problem with it.

To add it, open your project from the sidebar and click the pencil next to Instructions:

Paste the line and click Save instructions:

In Claude Code, put the line in your CLAUDE.md instead.

Leave it out of projects for long analysis, where a later step can show a mistake in an earlier one.

One habit to copy

When a long task ends, Claude’s summary says what it did, what it found and what it needs from you. Read the last part first: a decision it left open, or a change it wants you to approve.

You don’t have to change anything for this, since the summary already has a default format. But if you want what needs you to stand out, you can end your prompt with:

End with three headings: Blocked on me, Changed, Found.

If you want it on every task, put the line in your project’s instructions instead.

Subscribe now

3. Pro: instructions that run without you

At this level, you stop writing prompts and start writing the instructions Claude reads every time.

I run dozens of routines with Claude Code every night, and Anthropic just changed the rules for writing them.

Here’s what I’d change first:

1. Tell it which stops you want: Claude now keeps you posted as it works, and sometimes it stops to report instead of going on. Put a short rule in your CLAUDE.md file.

Copy this prompt:

When a step doesn’t need my input, keep going. Put status notes in the same message as your next action.

Stop and ask only when you can’t continue without me, or before anything destructive: deleting data, force-pushing, or changing anything outside this repository.

When you tell Claude to keep going, it stops less often. So keep one stop for anything risky or hard to undo. That’s what the last line of the rule does.

2. Split big work across subagents: for an audit or a change across many files, ask Claude to hand each piece to its own subagent and check each result before accepting it.

Here’s a template, based on Anthropic’s example:

Check every [item] in [the folder or list] for [what you’re looking for]. Give each [item] to its own subagent. When a subagent reports back, check its evidence before you accept it.

3. Trim your old instructions: Anthropic removed over 80% of Claude Code’s own system prompt for Claude 5 models, with no measurable loss on its coding tests. Your CLAUDE.md files and skills probably have the same problem: rules written for older models that now get in the way. Run /doctor in Claude Code, and it’ll help you trim them.

4. On Fable, tell it the user isn’t watching: left alone on a long job, Fable sometimes stops to ask permission for something your first message already covered. Anthropic’s fix starts with one sentence that carries most of the effect:

You are operating autonomously. The user is not watching in real time and cannot answer questions mid-task, so asking ‘Want me to…?’ or ‘Shall I…?’ will block the work.

The full version is in my Claude Fable 5 guide.

4. The anatomy of a Claude prompt

Here’s everything in this guide in one prompt.

You’ll rarely need every line. A quick question needs one. Use the parts your task needs.

Here’s an example:

You’re helping me write [a one-page summary of our Q3 report / a client email / a project plan].

It’s for [my manager / my team / a client], so keep it [short and plain / formal / friendly].

Only use numbers from the report. My manager checks every figure.

Match the structure of last quarter’s summary.

I’ve attached [the report / last quarter’s summary / a dashboard screenshot].

Give me [a one-page Word doc / a spreadsheet with one row per client / a 5-slide deck].

Don’t use [buzzwords / bold text / a recap].

Mark anything you couldn’t confirm, and say where you looked.

Done means: [one page, three wins, one problem and one number to watch].

Stop and ask me only if [the report and the dashboard disagree on a number].

Before you start, ask me any questions you have.

Three of these lines stay the same from task to task: the defaults to avoid, marking what Claude couldn’t confirm, and asking questions first.

Put them in your project’s instructions once, and your prompts get shorter.

One more thing

Don’t add every line to your next prompt.

Pick the level you’re at, add one line, and see what changes in the answer. Then add the next one. A line earns its place in your prompt when you can see what it changed.

If you know someone who still uses “think step by step” in their prompts, send them this guide.

You now know how to prompt Claude. Paid subscribers get a deep-dive guide every week, plus free access to my video courses. If this one was useful, join them 👇

Subscribe now

For more, check out the docs.

A few days before this session, she asked me:

“I use Claude chat every day now. What else can Claude do?”

She knows the chat box. Now it’s time for the part of Claude that does the work: Cowork.

Here’s the main difference: Claude Chat gives you text, the way ChatGPT did in 2023. You ask, it answers, and the work is still yours. Cowork does the work: it builds reports, runs workflows and handles the boring clicks for you.

We had a 38-minute session, and I kept it to the three features I use in every automation I’ve built.

This guide walks you through each one step by step, with a screenshot for every click, so you can follow along on your own computer. The complete video is at the end.

In this guide, you’ll learn:

• The 3 Cowork features that every Claude user must know

• Claude in Chrome: Claude does the clicks for you

• Skills: three ways to get a skill, and the one I use 80 to 90% of the time

• Scheduled tasks: how to make Claude run a skill every night while you sleep

• How to go from Claude beginner to pro

Before we start, here’s a quick video on what this beginner series with Diana is about:

Subscribe now

The 3 Cowork features that matter most

You can skip almost everything in Cowork except three things:

• Claude in Chrome: Claude does the clicks for you.

• Skills: Claude remembers how it did them.

• Scheduled tasks: Claude does them again while you sleep.

They show up in every automation I’ve built. They’re the foundation every Cowork user needs.

This guide focuses on these three, one section each, in the order you’d use them.

Note: Anthropic recently merged chat and Cowork into one Claude. The Chat and Cowork switch is gone: you type in one box, and Claude figures out what each request needs. It’s rolling out to Pro and Max plans first, with Team and Free plans next. Nothing changes for this guide.

1. Claude in Chrome: Claude does the clicks

Claude in Chrome is an extension for Chrome, the browser most of us already use every day. It lets Claude use Chrome for you. It clicks, types and downloads the way you would.

Picture scheduling a social media post. You click New, then Post, pick LinkedIn or Instagram, type the text and attach a PDF. Every step is a click you make with your mouse.

Doing it once is fine. Doing it every day is work that deserves to be automated.

Here’s how to set it up:

1. Search “Claude in Chrome” on Google and open the Chrome Web Store page.

2. Click Add to Chrome, then Add extension.

3. Log in to your Claude account and click Authorize.

4. Click the puzzle icon at the top right of Chrome and pin Claude, so it’s always one click away.

In the session with Diana, we tried Claude in Chrome on something simple: saving an article as a PDF (if you want a more robust use case, I’ll link to one at the end).

Let’s do it together!

Open any of my guides on artificialcorner.com (or try it on this one), click the Claude icon to open the sidebar, and type:

download this article as a PDF file

Doing this yourself means pressing Ctrl+S (Cmd+S on a Mac) and saving the page as a PDF.

Claude does it on its own and hands you the file.

One habit I’d copy: I use the sidebar only for the first test. It answers one question: “Can Claude do this?”

Once it can, I move to the Claude app, which keeps everything organized, and write something like:

use Claude in Chrome to download this article as a PDF: [link]

The sidebar gets messy fast. The app is where the real automations live.

From the next section on, we’ll mostly work in the Claude app.

Subscribe now

2. Skills: Claude remembers how

A skill is a saved recipe. It holds the steps Claude followed to finish a task successfully, so it can follow them again next time.

That also tells you when to make one: only for tasks you repeat, never for something you’ll do once. For example, I save every article I publish as a PDF, every night, so that task got its own skill.

Here are three ways to get a Claude skill.

Way 1: Turn a run that worked into a skill. This is what I do 80 to 90% of the time. When Claude finishes a task the way you wanted, type “turn this into a skill”, then click Save.

That’s what we did with the PDF from the previous section: once Claude in Chrome had saved the article, Diana typed the prompt below in the same chat:

turn this into a skill

The next run follows the same path, without the mistakes from the first try

Way 2: Record a skill. Click the + button and choose Record a skill. Then you do the task on screen while Claude watches, like showing a new colleague how you do it.

Here’s me recording one. My gym videos land on my Mac with generic names (IMG_4521, IMG_4522, and so on), so I showed Claude how I rename them: watch the video, find the exercise and the weight, and count the reps.

Pro tip: keep the recording clean. If you make a mistake while recording, Claude copies the mistake too.

Way 3: Install one someone shared. Skills are shared as .zip files (like the ones in my best Claude skills). If someone sends you a folder, right-click it and choose Compress first. Then:

1. Click Customize in the left panel.

2. Open Skills.

3. Click Add at the top right.

4. Choose Upload skill and pick the .zip file.

Note: There’s also a Create a skill option. I don’t recommend starting there. You need to explore a task first, and a skill written before a run that worked starts with all its mistakes built in.

3. Scheduled tasks: Claude does it while you sleep

A skill only runs when you trigger it by hand, unless you give it a schedule.

That’s what a scheduled task does: it runs the skill for you at whatever time you choose, even while you sleep.

The easiest way to create one is to ask Claude in a chat:

1. Open a new chat in the Claude app (select Cowork if you still have the Chat and Cowork switch).

2. Type / to open your skills and pick the one you want (mine is article-to-pdf).

3. Add when it should run: “schedule the /article-to-pdf to run every day at 2 AM”.

Claude may ask you a few questions back. Ours asked whether to download the latest article or a fixed link. That’s a good sign: every answer makes the task more precise.

Once it’s created, here’s how to find it and see what Claude set up:

1. Click Scheduled in the left panel.

2. Click your task to open it.

3. Click the pencil icon at the top right, next to Run now.

You’ll see everything Claude set up: the name, the instructions, how often it runs and what it’s allowed to do.

You can change any of it by hand, but I’d ask Claude to edit it instead. Long instructions depend on each other, and Claude updates everything a change affects.

4. How to go from Claude beginner to pro

Chrome, skills and scheduled tasks are the basics. Everything else in Cowork builds on them.

The rest of this guide isn’t more theory. It’s what to do this week, in the order I’d do it, and each step comes with a guide that walks you through it.

Below the line:

• The use case that puts all three pieces to work. Our job-hunting series: Claude finds the job posts, a skill scores them against your CV, and a scheduled task runs it every morning.

• The Cowork guides worth reading next. Your first automation, the files that tell Claude about you and how you write, and your workspace setup, in the order I’d read them.

• The full video session with Diana. Unedited, 38 minutes: Diana installs Claude in Chrome, saves her first skill and schedules it, including every moment a button wasn’t where we expected. If you get stuck on a step above, you’ll find it there.

The rest of this guide is for paid subscribers. Support us by becoming one.

Subscribe now

5. What to do this week

Read more

Today’s guide was written by Diana, who’s part of the Artificial Corner team. I’m bringing in writers like her so you get more tested guides every week than I could ever write alone.

Paid subscriptions are what make this possible.

They’re how I can bring in people who know a topic better than I do. Paid subscribers get full access to all my guides, my video courses ($200+ value) and these perks.

If you’d like to support the team, consider becoming a paid subscriber 👇

Subscribe now

OpenAI recently released ChatGPT Images 2.5. According to OpenAI, it’s up to 50% faster, better at keeping the people in your photos looking like themselves, and better at editing only what you ask.

So I put it next to Nano Banana 2 (Google’s image model, inside Gemini) and gave both the same prompts.

Overall, ChatGPT made the best-looking pictures. It also refused twice, and it still took close to a minute per image. Nano Banana was faster.

Here’s what I found, test by test, and how to pick the right one:

• Where to find both tools (both work on the free plan)

• Eight head-to-head tests, and which model won each one

• How to test both with your own prompt in a few minutes

• Which one to pick for your next image

Click here to get my FREE AI courses (20+ lessons)

Subscribe now

How to access Nano Banana and ChatGPT Images

Both live inside apps you may already use, and both work on the free plan.

Nano Banana 2 is inside Gemini (the AI app from Google):

1. Go to gemini.google.com.

2. Click on “Tools” → “Create image”.

3. Make sure to use the “Thinking” model.

Free users get the “Fast” and “Thinking” models. The “Pro” model needs a paid Google AI plan.

ChatGPT Images 2.5 is inside ChatGPT:

1. Go to chatgpt.com.

2. Click on “More” → “Images” (or skip this and describe the picture in a normal chat).

3. Type your prompt and wait. A detailed image can take a few minutes.

The free plan gives you only a few images per day (OpenAI does not publish the exact number). “Images with thinking” is for Plus, Pro and Business plans.

Here’s the first test. I gave both models the same selfie and the same prompt for a magazine cover:

My favorite is the ChatGPT one on the left, and it won’t be the last time ChatGPT wins on looks.

Subscribe now

ChatGPT vs Nano Banana

I gave both the same source images and the same prompt, every time. Here’s what came back.

Cut a product out of a messy background

Prompt: High-end commercial studio product photography of the isolated core product, centered on a pure, seamless solid white background, 50mm lens, sharp focus, shot from a slight three-quarter angle at eye level. Soft balanced studio lighting via dual softboxes, clean reflections, subtle natural drop contact shadow grounding the base. Perfectly preserved original textures, labels, typography, logos, and materials with zero distortion. Ultra-detailed, photorealistic, 8k, commercial catalog grade, no artifacts, no extra objects, no clutter.

Both models did this well. Clean edges, small details kept, nothing strange around the product. A tie.

Infographic

Educators can use this feature to turn photos into photo-based infographics with a single prompt. I gave it an image of the Golden Gate Bridge, and it turned it into a fantastic infographic.

Prompt: Turn this picture into a detailed educational photo-based infographic explaining important details on the dimensions and materials

Both did a good job here too. Clear layout, readable text, and the bridge still looked like the bridge. Another tie.

Combine images for high-quality campaigns

Say you have the elements below, but don’t want to spend time putting them together in Photoshop.

Prompt: Put these five people and this dog into a single image, they should fit into a stunning award-winning shot in the style of a fashion editorial. The identity of all five people and their attire and the dog must stay consistent throughout but they can and should be seen from different angles and distances, as is most natural and suitable to the scene. Make the colour and lighting look natural on them all, they look like they naturally fit into this fashion show.

ChatGPT won this one. Its image looked like a real photo. The Nano Banana 2 version looked more like pieces glued together.

Making a collage

Prompt: Create a collage from these two photos - the first one is from the internet and the second is my own original - and add a caption at the top.

Here ChatGPT was the clear favorite. It understood the task right away. I tried many times to get a nice collage out of Nano Banana 2, and nothing looked good.

Nothing on the web can be taken at face value anymore

Prompt: Change the figures in the bill so that the total comes to $20.

Here Nano Banana 2 was the clear favorite. It did the job on the first try. ChatGPT refused. I tried several times, I told it the image was for marketing, then for education. Still no.

Subscribe now

Creating a postcard

Prompt: turn the style of image 1 into the style of image 2

ChatGPT understood the design and the idea better. The result really looked like a postcard, not just two pictures mixed together.

From concept to reality

Prompt: Turn this into photorealistic image

A sketch in, a real-looking product out. Nano Banana 2 did it well on the first try. ChatGPT did not want to do it at all.

Click here to get my FREE Claude course (20+ lessons)

How to test both with your own prompt

My results are a starting point. Your prompts are different from mine, so run your own small test.

You don’t need any setup, and it takes a few minutes.

1. Open both tools in two browser tabs.

2. Write one prompt for an image you really need. Say what is in the picture, where it is, the light, and how close the camera is.

3. Paste the same prompt into both tabs. Don’t change a word.

4. Make two images in each tool, not one. One lucky result tells you nothing.

5. Zoom in. Check hands, text, and small parts like zippers and buttons.

6. Pick the winner.

Here is a prompt shape you can copy:

A [product or person] in [place].
[Light: soft morning light / studio light / evening].
[Camera: close-up / half body / full scene].
[Style: realistic photo / flat illustration / poster].
Text on the image: "[your exact words]".

Filling in every bracket yourself takes time. The easier way is to give ChatGPT or Gemini your rough idea and let it write the prompt for you.

Here’s an example of what comes back:

Prompt: A minimalist glass serum bottle on a wet marble surface in a sunlit bathroom. Soft morning light casting subtle water reflections. Close-up shot with shallow depth of field. Realistic photo, high detail. Text on the image: “Pure Radiance”.

Pro tip: if your image includes numbers or a chart, write every number in the prompt. Both models invent the missing ones.

Cost and speed

For one image, you will not feel a difference. The gap shows up when you make many images:

• Price: in the apps, you don’t pay per image. Both models are available on the free plan, with a daily limit. A paid plan gives you more images per day.

• Speed: Nano Banana 2 gave me an image in about 10 seconds. ChatGPT Images 2.5 needed around 1 to 2 minutes.

Which one to pick

Two things stood out after all eight tests. Nano Banana 2 says yes more often. But when ChatGPT does make the picture, its best results have the better quality.

Use ChatGPT Images 2.5 when:

• The image has text (covers, menus, ads, posters)

• Every detail of your prompt has to be followed

• You want a close, photo-style shot of one product or one person

• You edit a photo and the original has to stay the same

Use Nano Banana 2 when:

• You need the image fast

• You want to change the numbers in an image, like the total on a receipt

• You want to turn a cartoon or a sketch into a realistic photo

• ChatGPT refuses your request

One more thing

If you have access to both tools, use them both.

People online talk about these two as rivals. For a beginner, they are closer to two coworkers with different skills. I now run the same prompt in both, then keep the better picture.

Try it with an image you need this week.

If this one was useful, consider becoming a paid subscriber 👇

Subscribe now

Join my newsletter with 80K+ people to master AI in 2026

Let’s connect! You can find me on LinkedIn, Instagram and X

Most of it is briefing. I explain a task the way I’d explain it to a colleague: the context, the goal, what to avoid. And I say it out loud instead of typing it. Since May, that’s 1,174 recordings and almost 7 hours of me speaking my second language, the way I’d speak in a real conversation.

For months, all of it disappeared the moment Claude answered. Hours of speaking practice, and nobody ever corrected a word.

So I built a simple setup.

Every night, Claude reads what I dictated that day, checks it against my Language DNA Profile (my language preferences, my level, the mistakes I keep making) and writes a report made for me. It has run every night since August.

You can do the same with Spanish or any other language, even if you never talk to Claude in the language you’re learning. Dictate a few lines of your journal, a text to a friend, or tomorrow’s to-do list. Your dictation app keeps it, and Claude corrects it at night.

In this guide, you’ll learn:

• Why everything you dictate is speaking practice

• What you need: a dictation app that saves what you say

• How to set up a nightly tutor that corrects you while you sleep, step-by-step

• How to turn a pile of reports into a mistake notebook you’ll actually open

Subscribe now

1. Everything you dictate is speaking practice

When you dictate, you explain the whole thing, the way you’d explain it to a colleague: what you want, why, and what to avoid. That’s everyday speech, the kind a language teacher wants to hear.

Here’s who this works for:

• You already dictate in a language that isn’t your first, to an app, in messages or in your notes. The practice is already happening. You just need to keep what you said.

• You’re learning a language you don’t use every day. Dictate something you’d write anyway, but in Spanish: a few lines of your journal, a text to a friend, tomorrow’s to-do list, or one question to Claude. Your dictation app keeps it either way.

No partner, no class, and nobody watching you struggle.

The feedback works like a teacher with a red pen on your diary. You were writing in it anyway. Now someone corrects it every night.

A few minutes of dictation you’d do anyway beat half an hour of Duolingo you keep skipping.

2. What you need: a dictation app that saves what you say

Dictation turns your voice into text wherever your cursor is.

There are plenty of AI dictation apps, and I compared the best ones for talking to Claude in my dictation guide.

Here are the most popular dictation apps:

• Superwhisper (Mac, Windows): keeps the audio and the transcript, and it never deletes them on its own

• Wispr Flow (Mac, Windows): keeps your transcripts, plus the audio for a limited time

• VoiceInk (Mac): keeps the audio and the transcript of every dictation (this is what I use)

• Built-in macOS or Windows dictation: keeps nothing, so it doesn’t work for this

Other dictation apps work with this method too as long as they keep a history of your dictation.

This is my VoiceInk history from today. Every line is something I said to Claude while working.

Note: in my dictations up there you can see how I restart sentences and repeat words all the time: “how I’m how this method helps me learn a language”, “because it translates sorry it transcribes”. Everyone does that, in any language, including in their first one. The prompt in Step 3 tells Claude to ignore that, so it never counts as a mistake.

Subscribe now

3. How to set it up

Here’s what to do, in order:

Step 1: Turn off the AI cleanup

Most dictation apps can fix your grammar before you see the text. That’s great for emails. For this, it erases the exact mistakes you want feedback on.

Here’s where to check:

Wispr Flow is the most popular of these apps, so let’s start there.

Option 1, no cleanup at all: go to Style → Auto Cleanup and pick None. Flow then “transcribes exactly what you said, including mistakes”.

Option 2, keep your cleanup: if you like how Flow polishes your messages, leave it on. When a dictation is worth practicing with, click the three dots next to it in Home, then “Undo AI edit”. You get back exactly what you said (you’ll have to do this manually every time though)

VoiceInk: AI enhancement is off by default. Leave it off.

Superwhisper and other dictation tools: use a mode with no AI step

Step 2: Build your Language DNA Profile

The profile is one file about you: your first language, your level, your goals, and the mistakes you already know you make. Claude reads it every night before it corrects you, so the feedback fits your level and your first language.

To build it, use the 50-question interview from my guide. To do it faster, answer by dictating in your first language, and at the end Claude writes the file for you. Save it as language-dna-profile.md: you’ll put it in a folder in Step 3.

Step 3: Set up your nightly tutor

I set up Claude to give me language feedback on everything I say during the day. It works while I sleep, and the report is waiting for me when I wake up.

I’ve used it since August. That’s 29 nightly reports so far. I’ll also show you where to click, step-by-step, so yours starts tonight.

The setup below is one piece of a bigger system.

The whole thing, including my video course on learning a language with AI, comes with the paid subscription, along with other perks and full access to new guides every week.

Subscribe now

Read more

The one you probably haven’t used is Fable, the one above them.

And Fable isn’t even the top. Anthropic has one more model, called Mythos, and you can’t use it. In testing, it beat nearly every human expert at finding ways to break into software, and found thousands of serious holes in every major operating system and browser. Anthropic decided that’s too much power to give to anyone, so Mythos only goes to a handful of partners pointed at defending critical systems.

That makes Fable the most capable model the rest of us can get.

Same family, one step below the one they locked up, at double the price of Opus. I kept Opus as my daily driver and use Fable for specific jobs.

In this guide, you’ll learn:

• Where Fable lives

• What Anthropic says Fable is for, and when to ignore it

• The jobs I actually use it for

• How to prompt Fable 5.1

• And whether your task deserves it

Subscribe now

1. Where Fable 5 lives

Here’s where to switch it on:

In the Claude app:

1. Open a new chat

2. Click the model name at the bottom right of the message box

3. Pick Fable 5.1

Hover over it and Anthropic tells you the catch itself:

Fable is the most capable model and draws down usage much faster than Opus

In that same menu, click Effort:

Five levels: Low, Medium, High, Extra, Max. Medium is the default in the Claude app, and Max carries its own warning: “3× or more usage.”

Everything in this guide ran at High.

In Claude Code:

1. Type /model

2. Pick Fable

3. Press Enter to make it your default, or s to use it for this session only

Note the line at the bottom of that screen: High effort (default).

Claude Code sets it high for you. Adjust it with the arrow keys, or skip the menu and launch the session using this command:

claude --model fable --effort high

Now, switching Fable on is the easy part. Knowing when to leave it on is the actual skill, and I pick Fable job by job.

The rest of this guide is about which jobs and how to prompt Fable 5.1

If you’re not a paid subscriber yet, now’s the time to join hundreds of members with full access to all our guides :)

Subscribe now

Read more

For this series, I have a student: Diana. She knows a little bit of ChatGPT and Claude. When I asked her what she uses AI for, she said:

“Normally I use it for studying, work, Excel, googling.”

That’s most people.

What she has never touched: Claude Cowork, ChatGPT Work, Claude Code, Codex. Almost zero experience there. That’s exactly why she’s the perfect student, and if that describes you too, this guide is written for you.

We had a 37-minute session: the very beginning, the part nobody records. The complete video is at the end.

Here’s a decision I made for this first guide in the series: I’ll start with the foundation.

If you learn the foundation, you can pick up Claude, ChatGPT, or whatever goes trending next, in an afternoon. These companies are rivals, so almost everything one has, the other has under a different name.

Learn the shape once and the names stop mattering.

In this guide, you’ll learn:

• What ChatGPT and Claude actually are (and the one word that scares people off)

• Why “ChatGPT vs Claude” was never the right question

• Which model is right for you, and why you probably don’t need the best one

• The three levels Claude comes in

• How to talk to Claude, what it’s good and bad at, and what to try this week

• Your first step to go from AI beginner to pro

Subscribe now

1. What Claude and ChatGPT actually are

Remember the first time you opened ChatGPT.com? One box. You typed in it. You used it like Google, or for translation, or to get some facts.

That’s what both ChatGPT and Claude were when they launched, and it’s still what the chat window is today. The simplest way I can put it: a smarter Google. Everything Google gave you, plus it answers you directly instead of handing you ten links.

On top of that, it could write code.

I know “code” sounds like a word for programmers, and you’re about to skip this paragraph. Don’t. This is the single most important thing in the guide.

When I say code, I don’t mean programming.

I mean things that happen on your computer. Code is how a report gets built without you building it. Code is how three hundred photos get renamed while you make coffee. Code is how a system runs every night whether or not you remember to run it.

So when I say these tools can write code, what I’m actually telling you is: you can now do things on your computer that were not possible for you before.

Not for programmers. For you.

That’s the whole promise, and everything in the rest of this guide is a bigger version of it.

2. ChatGPT vs Claude was never the right question

Halfway through the call Diana asked the question everyone asks:

“Which is better, ChatGPT or Claude?”

Wrong question, and I understand completely why everyone asks it. The right question is: which model do I need for this task?

I’ll get to models in the next section. First, why the comparison itself is the wrong question.

I’ve written ChatGPT vs Claude comparisons. They got a lot of readers. And I want to be honest with you about them.

Every comparison I’ve published was true the day I published it. Six months ago I wrote that Claude was clearly better at writing and clearly better at coding. That was true.

Today it isn’t.

GPT-6 Astra came out on September 3 and it beats Claude at many things.

Here’s the problem. You’re reading this weeks or months after I wrote it. If I hand you a table of who’s winning, you’ll believe it, and by then it’s wrong. Not because I lied. Because these two companies leapfrog each other every few months. Claude pulls ahead. Then ChatGPT ships a model and pulls ahead. Then it flips again.

A Claude vs ChatGPT comparison is true today and wrong in a month.

What is worth telling you is something different and much more stable: what one has that the other simply doesn’t.

• Claude doesn’t generate images. ChatGPT does, Gemini does, Claude doesn’t

• Claude’s voice mode is weak. It exists, but it’s nowhere near ChatGPT’s. I practise languages with voice mode (Spanish, French) and for that I use ChatGPT.

• Claude never had video generation. ChatGPT had Sora, and even that is gone now

Those gaps have held since 2023 and they’re still true now (image, voice and video aren’t Claude’s priority)

That’s a much better thing to know than a leaderboard, because it doesn’t flip next month.

Everything else? They mostly have the same features. One will be stronger than the other at any given moment, and then they’ll swap.

Subscribe now

3. A better question: which model is right for my task?

Open Claude and click the model name in the corner. You’ll see a list.

Think of it as a ladder:

• Haiku is the small, fast one. Quick questions.

• Sonnet is your everyday model. Emails, drafts, translation, learning something. Most of what you do, most days.

• Opus is the serious one. Building a system, setting up automations, anything where reliability matters. When something involves real work rather than a quick answer, this is where I go.

• Fable is the top of the range. High-stakes work: money involved, or a mistake would be expensive.

My own split, roughly: Opus for most of it, Sonnet for everyday things, Fable rarely.

ChatGPT has the same ladder under different names. Terra and Luna are the everyday ones (Terra is the balanced one, Luna the fast one), and GPT-6 Astra sits at the top, the equivalent of Fable.

Now the question I actually want you to ask yourself: do you really need the top model? Do you need Fable? Do you need GPT-6?

I don’t, most of the time. Almost nobody reading this does. Powerful also means expensive, and for everyday work you’d be paying for headroom you never touch. You need it when you’re building something complex, or when a mistake costs money. Everything else, Sonnet or Terra will do, and you’ll never notice.

So the answer to Diana’s question is: neither is “best.” Best is a model, not a company, and the model you need for most of your work is the cheap one.

4. How to get them

Both work in the browser with a free account:

• Claude: claude.ai. The desktop app, which you’ll need for level 2 and 3, is at claude.ai/download.

• ChatGPT: chatgpt.com. Desktop app at openai.com/chatgpt/download.

Both have a free plan. Both free plans are very limited. You won’t see what these tools can really do on them.

The paid tiers are the same story on both sides:

• $20 a month for real access with a smaller allowance (enough for beginners)

• $100 for a bigger one

• $200 for the most generous (that’s where I am right now).

Every plan has a limit, including the expensive ones.

I’m on the $200 plan and I still hit the limit some weeks.

Not from chatting. From the heavy tools, the ones where Claude takes over your screen, clicks around, and takes a screenshot at every step to see what it did. Each of those screenshots costs usage.

If you only chat, you’ll never see the wall. If you start automating, you will, and I had to learn how to cut my usage before I stopped hitting it.

Subscribe now

5. The three levels

Every one of these tools comes in three levels. Once you see the ladder, the whole product line stops being confusing.

Level 1: Chat

The box. The smart Google that can code. It exists on day one of every one of these products, it’s on your phone, it’s what everyone means when they say they “use ChatGPT.”

Chat talks. That’s the thing to hold on to. It can produce anything, including code, but it hands it to you. What you do with it is your problem.

Level 2: Cowork / ChatGPT Work

Chat talks. Cowork works.

Here’s the difference in one example. In 2022, if you wanted to automate an Excel report, you’d ask ChatGPT to write you a script, and then you’d be sitting there with a script, wondering what to do with it.

With Cowork, you skip all of that. You don’t see “code” at all. You only say: automate this report for me. Cowork goes to the folder where the report lives, and it does it. It might write a script along the way. That’s its business, not yours.

Three things I’ve actually used Cowork for on week 1, none of which need technical knowledge:

• Renaming a trip’s photos. Everything comes off the phone as IMG_1234. Point Cowork at the folder and it looks at the photos and renames them by what’s in them: the restaurant, park, beach.

• Tracking subscriptions. Give it your bank statements. It finds Netflix, HBO, all of them, and tells you what you’re paying every month.

• Automating a report and creating interactive dashboards (aka artifacts)

Those are small on purpose, so you can see the shape.

I run a real part of my business on Cowork, and the systems are much bigger than this.

I mentioned “Cowork“ because it’s the tool I’ve been using mostly, but the same applies to ChatGPT Work. It’s the same thing: same idea, same way of working.

Level 3: Claude Code / Codex

The name is intimidating, but it shouldn’t be. You don’t need to learn to code to use it.

Level three is more powerful than level two for one reason: access.

Cowork is sandboxed. That word means it runs inside a safe box. It can still touch your files, but it’s fenced in, and it can’t reach everything. If you’re new and you’re trying automations for the first time, that fence is protecting you.

Claude Code isn’t fenced in the same way. It has access to your whole computer. That’s the scary part, and it’s also exactly why it’s more powerful.

Two simple examples of why I use Claude Code more than Cowork.

Reaching tools on the internet. I wanted to download videos from YouTube. In Cowork, there was no clean way. It told me it couldn’t, and the workarounds were slow and expensive. In Claude Code it took no time at all, because it could reach a tool somebody published on GitHub called yt-dlp, connect to it, and do the job. Cowork couldn’t reach out to that. Code could.

Changing many things at once. I have tasks that run on a schedule, every night. In Cowork, each one is an island: to change something in ten of them, I open ten tasks and change it ten times. In Claude Code I say “change this in all of them” and it’s done in one shot.

Diana’s question at this point was the one you might be wondering now:

“Does ChatGPT have something like this?”

Yes. It’s called Codex. Finding it on her screen took her a few tries, though.

The ladder, in one line: chat, then Cowork, then Code on the Claude side. Chat, then work, then Codex on the ChatGPT side.

Easiest to hardest, safest to most powerful, in that order.

Which also answers the other question Diana asked:

“Which one is easier to use, ChatGPT or Claude?”

Neither. They’re the same. The real answer is chat, then Cowork, then Code, and it’s identical on both sides. Start at chat if you’re absolutely new, otherwise start with Cowork or ChatGPT Work.

6. How to talk to these tools

Three habits carry almost all of it. Each one comes with a line you can copy.

1. Tell it who it’s for and what happens next

Weak:

Write an email to a client.

Strong:

Write a reply to a client who missed our call on Tuesday. Friendly but firm. She needs to pick a new date this week. Four sentences.

The details you’d give a new assistant are the details it needs. Who, what happened, what you want to happen, how long.

2. Treat it as a conversation, not a form

Nobody writes the perfect prompt first time. Type your prompt, press enter, see what comes back, then steer:

• “Shorter.”

• “Warmer. Add a line about the deadline.”

• “Good. Now the same thing for my boss.”

Start a new chat when the answers start feeling off, the chat is full and the model is losing the thread. But don’t just open a new chat and start over. First, ask for a hand-off:

Summarize what we’ve decided so far and what’s still open, so I can paste it into a new chat.

Copy the answer, open a new chat, paste it in, and keep going.

3. Give it an example

Don’t describe what you want. Show it. Paste something you wrote and liked, then:

Write a new one like this. Same tone, same length.

Once it learned how you write, the next responses in that chat will follow the same style.

7. What they’re good at, and what they’re bad at

Good at:

• Writing and rewriting, in your voice if you feed it examples.

• Long documents. Drop in a hundred-page PDF and ask questions about it. Ask for page references so you can check the answer instead of trusting it.

• Making files. Both will build you a spreadsheet with real formulas, a Word document, a slide deck or a PDF.

• Thinking a decision through with you. Genuinely useful, with one condition I’ll come back to in section 10.

• Your files and your computer. Levels two and three. The reason the rest of this series exists.

Bad at:

• Vague requests. You get what you give.

• Telling you it doesn’t know. Its default is to answer. Treat it as a very fast colleague, not as a source of truth.

The rule underneath all of it: it will always give you an answer, and that is not the same as the answer being right.

8. Your first step to go from AI beginner to pro

The rest of this guide is for paid subscribers.

The rest of this isn’t more theory. It’s the things to do this week. They come with guides that walk you through it step by step.

Below the line:

• Six things to do this week, ranked from easy to real. Three you can do in the chat box today, three that use Claude Cowork. Each one links to a complete guide so you’re never left staring at a blank prompt.

• The one thing I’d set up first for better decision-making with AI. There's one thing you can set up that makes Claude noticeably better. It changes the answers you get back, and it's the difference between AI helping you decide something and AI just telling you what you want to hear.

• The full video session with Diana. Unedited. Every wrong click, every “where is that button,” nothing cut.

Subscribe now

9. What to do this week

Read more

Click here to get my FREE AI courses (20+ lessons)

OpenAI released GPT-6 Astra on September 3.

Greg Brockman, OpenAI’s president, called Astra a generational leap and said it could eventually be seen as the arrival of artificial general intelligence (AGI), which is an AI that can handle any job you’d give a capable person, instead of being very good at a handful of things. It’s the finish line the whole industry keeps pointing at, and nobody agrees on how we’d know we’d crossed it.

I did three things.

I watched people putting it through real work.
I read OpenAI’s own guidance on using this model.
I gave it a job with real money in it: the same one I handed Fable last week.

It surprised me in both directions.

In this guide, you’ll learn:

• Where Astra lives

• Where Astra wins and loses (and what else came out of it)

• What happened when I gave Astra the job I gave Claude Fable last week

• How to prompt GPT-6 Astra

• Whether the $20 plan is enough for Astra

• And whether you actually need Astra

Subscribe now

A note for Claude users

I’ve spent most of this year inside Claude.

There’s a plain reason for that. For the first half of this year, Claude was the better tool and it wasn’t close. Cowork already existed, and ChatGPT had nothing like it. Their equivalent, ChatGPT Work, arrived months later. If you wanted to hand a computer a real job and walk away from it, there was one serious option.

That gap has narrowed. That’s why I’m writing this one. But this isn’t a switch-to-ChatGPT piece.

I never cancelled ChatGPT. I downgraded it to the $20 plan back in January and left it there, because two things on it stayed better than Claude’s: voice mode, which has never really been close, and image generation. Meanwhile I moved up to Claude’s $100 plan, and then the $200 one.

Here’s what I didn’t expect: the best thing I got from a week of testing Astra was a better way to use Claude.

So read this as a guide to getting more out of whichever one you already pay for, with a clear answer at the end about whether Astra is worth adding.

1. Where Astra lives

There are two places to switch it on.

In ChatGPT Work:

1. Download ChatGPT for desktop. Open ChatGPT and click Work

2. Click the model name at the bottom of the message box

3. Pick GPT-6 Astra

4. Drag the effort slider next to it. Everything in this guide ran at High

In Codex:

1. Click ChatGPT in the top left corner and pick Codex from the dropdown

2. Click Select effort at the bottom right of the message box

3. Pick GPT-6 Astra and set the level, again High here

That effort slider decides how hard the model thinks before it answers, and it costs you accordingly.

ChatGPT Work is like Cowork, and Codex is similar to Claude Code.

Click here to get my FREE Claude course (20+ lessons)

2. Where Astra wins and loses (and what else came out of it)

Nate Herk published eight videos about Astra in six days and put his own money through it.

Here are the best insights:

1. He ran 15 real jobs against Claude Fable 5.1, and Astra won 10: a presentation deck, a sales page, his taxes, an audit of years of subscription emails, meeting analysis, video edits, a small app, browser tasks, a website, a year of YouTube stats.

Here are the numbers:

Astra came in cheaper, but took longer. OpenAI’s own docs explain the cheaper half: Astra costs more per word than the model before it, and uses far fewer words to finish a job, so the bill per task comes out lower.

Treat those as one careful person’s numbers rather than a benchmark, though.

2. Astra won every job where it had to operate something: his taxes, an audit across two email accounts going back years, opening Canva, painting a picture with the mouse and uploading course videos into a community platform.

3. Fable won the jobs where taste mattered: the presentation deck came back better structured and more on-brand. The sales page answered the questions a real buyer actually has. The interface it designed was the one Nate preferred to look at.

4. It made an entire YouTube video from a single prompt: it found the footage, wrote the script, used his voice clone and his avatar, built the edit, added music and sound effects, and rendered the file. About 50 minutes.

5. The caveat: Nate says plainly that for most knowledge work, Astra is overkill and he reaches for the cheaper model. He also warns it will crawl through your weekly usage limit.

6. Brainstorm the failures before you automate anything: before letting it trade with real money, he had Astra list the ways the setup could break. What happens on the first run when there’s no previous record. What if a run crashed after placing an order. What if two runs overlap. What if the machine goes offline.

I could get more insights after doing my own test.

Subscribe now

3. I gave Astra the job I gave Claude Fable last week

Last week I handed Claude Fable 5.1 a real job with money in it: replacing the spreadsheet one of my contributors uses to work out his share of paid subscriptions. His sheet had grown over two years and had workarounds in it I couldn’t verify.

This week I gave GPT-6 Astra the same spreadsheet and the same first prompt, word for word.

Same effort setting, and I rolled my own files back to the state Claude found them in, so neither model could see the other’s work.

Here’s what happened, and I’ve pulled out the part that applies to you rather than the part that’s only about me.

It read the file before it said anything. Then it interviewed me.

In my experience, the model that asks you one good question at the start is worth more than the model that’s cleverer and guesses. If you take one habit from this guide, make the thing interview you before it builds.

Then, while I was still answering, it carried on reading and started telling me what was wrong with the file: an old duplicate of the same sheet sitting inside it, repeated entries in one of the lists, formulas that could quietly change past totals.

It kept going. Astra found errors in my contributor’s spreadsheet that Fable never mentioned.

Nobody had looked at those details.

That’s the point. If you have a spreadsheet that decides money, payouts, invoices, commissions, hours, and you inherited it from someone or built it years ago, this is the first thing worth doing with one of these models. Before you ask it to build you something better, ask it what’s wrong with what you have.

Then I asked for a payout figure for every month, and it could only work from what I’d handed it.

That’s the difference. Claude went and fetched what it needed from my Substack account on its own. Astra never did. Not in ChatGPT Work, not in Codex, on my own machine with my own files sitting right there.

So I gave it the data, and it came back with a month-by-month table.

Fable and Astra landed about 5% apart on the total. Month by month, they barely matched.

That’s worth knowing if you ever check one of these against numbers you already have. The total tells you whether the method holds up. The breakdown tells you where it differs. Two sensible answers can agree on what you owe in a year and disagree about what you owe in March.

Both of them did the job, and did it well. Astra asked the better questions. Claude Fable was the only one that could go and get the answers.

If you’re not a paid subscriber yet, now’s the time to join hundreds of members with full access to all our guides :)

Subscribe now

4. How to prompt GPT-6 Astra

I read OpenAI’s own guide for this model so you don’t have to.

Here’s how to prompt GPT-6 Astra for better writing, better instruction following, the right amount of self-checking, and more (it works with Claude too):

1. Give it a banned-words list: Here’s the prompt OpenAI recommends to stop its own model sounding like an AI

Read more

Hope you had a great summer! Now that September’s here, I’m easing back into a regular writing rhythm. Part of that is a new project I want to share.

I’ve started a second Substack newsletter with Diana, focused on writing and content creation (touching on AI sometimes, but not exclusively). If that’s something you’re interested in, join here.

Now, let’s get into today’s guide.

We’ve been talking since we were kids. It comes naturally.

For many people, writing doesn’t.

I’ve talked to a lot of people who struggle with writing (from a simple email to a full report). They all say the same thing: they want to write, but it takes too long.

They sit in front of a blank page, they know the topic inside out, but somehow nothing comes out.

They’re good at what they do. They’re not good at writing (yet).

That gap is exactly why I use (and recommend) what I call the recording method.

Instead of forcing yourself to write, you talk. Whether it's a conversation with someone else or just you thinking out loud into a recorder app, you get a transcript. Hand the transcript to Claude, and it uses your voice.md file to turn it into a draft that sounds like you.

This is my favorite method to write with AI. I’ve used it for some of my best guides and you can use it to write long emails, essays, newsletters, reports, etc.

In this guide, you’ll learn:

• The tools you need

• How this method works

• What to do before and during the recording

• The exact workflow I use (prompt, .md file, etc)

• A walkthrough of my personal system (so you can build your own)

The tools you need

Here are the tools I use for this method (besides Claude):

• iPhone users: Voice Memos or call recording (phone app or FaceTime). You get a transcript automatically when the recording ends (no extra app needed).

• Android users: use your phone’s built-in voice recorder or call recorder

• Optional: ElevenLabs for cleaner transcripts.

One thing worth knowing: the default iPhone transcript is a little messy (around 90 words out of 100 come out right). I’ve never had a problem feeding that messy version to Claude, though. Claude is smart enough to catch that “Clod” in the transcript was supposed to be “Claude” based on context.

If you’re not a paid subscriber yet, now’s the time to join hundreds of members with full access to all our guides :)

Subscribe now

The two ways to do this

Read more

Click here to get my FREE Claude course (20+ lessons)

This is part 3 of a three-part Claude Cowork series. Don’t miss part 1 and part 2

Every morning at 5, while I’m still asleep, Claude checks the job board I pointed it at, reads every job posted in the last 24 hours, scores each one against my CV, and writes proposals for the ten best matches. By the time I’m up, they’re in a Google Drive folder named after today’s date, ranked 01 to 10.

I don’t start it. I don’t watch it. I read ten proposals and decide which ones to send.

That took three parts to build, and this is the one where you get all of it.

Part 1 used Claude in Chrome to scrape 161 job posts from Indeed, Glassdoor, and Upwork into a single Google Sheet. Part 2 built Head Hunter, a skill that reads your CV and scores every post against it, then drafts proposals for the best ones. Part 3, this guide, is the scheduled task that runs both of them every morning without you.

You don’t need to have read the first two. At the end, I hand over one folder with three prompts, one per part. Paste them in order and the whole chain is yours.

In this guide, you’ll learn:

• What a scheduled task is, in plain English, and where to find it in Cowork

• The one prompt that builds the entire Daily Job Hunt task, start to finish

• What the first run actually produced

Subscribe now

This was made in collaboration with . He 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, check out LearnAIWithMe

What a scheduled task is, in plain English

A scheduled task is a prompt Cowork runs for you on a schedule, in a fresh session, with your project folder attached. You write it once, pick a time, and it runs every day at that time.

To see yours, click on “Scheduled”.

Setting it up

When you click on “New task”, you can see two options.

• Create with Claude

• Set up automatically

Let’s click on Create with Claude.

The task we’ll create here will automate job hunting. It will find relevant jobs and write proposals for them automatically, without needing you, from start to finish.

If you’re not a paid subscriber yet, now’s the time to join hundreds of members with full access to all our guides :)

Subscribe now

First, copy the prompt below:

Create a scheduled task named Daily Job Hunt. Run it every morning at 5.

Here is what it does, start to finish:

1. Search Upwork for Data Science jobs posted in the last 24 hours.Pull the full post for each one: title, client, budget, experience level, description, and URL.

2. Open job-hunt/state/seen.json and skip every URL already in it.Those are jobs we already drafted or rejected.

3. Run the head-hunter skill on what is left. It reads cv-profile.md and scores every new post with its four criteria. Do not rebuild the CV profile, it already exists.

4. Draft proposals for the top ten. If fewer than ten clear the bar, draft what is there and say how many made it.

5. File the day into Google Drive. Inside my Job Hunt folder, create a folder named with today’s date. Upload each proposal as a Google Doc named NN - Client - Role, and the day’s scores as a Google Sheet named Scores YYYY-MM-DD.

6. Add every drafted and rejected URL to seen.json so tomorrow’s run skips them.

7. End with a short report: how many new posts came in, how many were skipped as seen, how many got drafted. If nothing new came in, report “nothing new today” and stop. Never redraft yesterday’s jobs.

Rules that never bend: never apply to anything, never message anyone, never sign in anywhere, never share a Drive file, never convert salary formats, never invent experience that is not in my CV.

After pasting this prompt, it’ll ask you to schedule this task.

I approved it. Claude created the task, then asked whether to run it right away instead of waiting for tomorrow morning. I said yes.

And it started working.

The first part is to collect data from Upwork. It will score the jobs and write proposals for the top 10, based on the criteria I defined.

You can click on Scheduled, then open your task and check the History section to see the process.

Here it is.

It scored the jobs, and now it is drafting the proposals. Check the progress bar, 3 of 5, at the top right.

After a few minutes, the task is finished.

Here is the report it sent. 50 results came back from Upwork. 30 were posted in the last 24 hours, 21 were worth opening, 13 scored above 45, and 10 got a proposal. The top score was 73, a senior data scientist role at $30 to 60 an hour.

Read the last paragraph too. My Mac went to sleep before it could write seen.json, and it told me so instead of pretending.

More on that at the end.

Let’s check Drive.

Here are all ten, in one Drive folder named after the date. The file names start with the rank, so 01 is the best match, and 10 is the weakest.

I opened number 10 on purpose, the weakest of the ten. It scored 50, and the skill says so in the first line.

Two pages. The score and its four parts, the post’s details as written, why it scored so, three CV points to lead with, the honest gap, and an opening paragraph I could send.

Look at the gap section. It says this is the weakest draft of the ten and tells me to send it only if I want a four-milestone commitment through June 2027. That is the part I would never get from reading 30 posts myself.

Click here to get my FREE Claude course (20+ lessons)

The chain, complete

If you did not read what we did in Part 1 and Part 2, that is fine. I put everything into one folder, and it walks you through the whole series.

👉 Here is the folder: Head Hunter

Download the three files, drop them into one empty folder with your CV as a PDF, and open that folder in Claude Cowork.

Then open START-HERE.md. It holds three prompts, one per part.

The first one sets up the project and runs the scrape from Part 1, the second runs Head Hunter from Part 2, and the third is the scheduled task you just saw.

Paste them in order, and the whole chain is yours in about an hour.

That is the whole chain. Claude in Chrome finds the posts, Head Hunter scores them against your CV, and the scheduled task runs both every morning while you sleep.

Ten proposals wait in Drive before your coffee, and the only thing left is reading them and clicking send.

One warning from my first run: the task runs on your computer, so keep the computer awake at 5. Mine went to sleep halfway, and Claude told me plainly which step failed and retried the next morning, but a closed lid still means a day without proposals.

Download the folder above, paste the three prompts in START-HERE.md, and you have the same thing running by tomorrow.

Join my newsletter with 80K+ people to master AI in 2026

Let’s connect! You can find me on LinkedIn, Instagram and X

Leave a comment

Click here to get my FREE Claude course (20+ lessons)

We run a business on Claude Code and Cowork.

started using Claude Code when the preview shipped in February 2025. I started using Cowork the day it launched.

There is a difference between using Claude and running a business on Claude. The first one forgives everything. The second one only holds up if the boring parts underneath it are right.

These are the lessons we learned running real work on top of them. Every one of these lessons is something dull we ignored for months, until it cost us a week, and then became the rule we never broke again.

If you’re new to Claude, check out my free Claude course

For more guides like this, support us by becoming a paid subscriber 👇

Subscribe now

Before you start: Claude Cowork or Code?

Both tools earn their place. This is not a question of which is better, but which one you are ready for.

I like to see it this way: Cowork is the first step. Claude Code is the last one.

If you are new to Claude, start with Cowork. You will get most of what people show off with Claude Code, with no terminal and zero code knowledge. I ran my entire publishing system on Cowork for months, and it held.

Claude Code is the full-power version. In Cowork, each job is its own island. In Claude Code, they are files sitting next to each other, so you (and Claude) can easily search them, reuse them, and fix all of them at once.

We learned 12 lessons after using them both. They come in two halves.

Lessons 1–6: For all Claude users

1. Separate the work from the system

2. Name folders, skills and tasks for what they hold

3. If you’ll edit it later, make it a variable

4. If you have many scheduled tasks, build a calendar

5. Give every job the same report shape

6. Tell Claude to delete what it removed

Lessons 7–12: For Claude Code users

7. Every line in your CLAUDE.md gets billed forever

8. A rule in CLAUDE.md is a suggestion — a hook is a guarantee

9. Learn just enough git

10. Delete half your MCP servers

11. Your best model should judge, not type

12. You are running one Claude. Run twelve.

Gencay helped me with the Claude Code lessons. He writes LearnAIWithMe, a newsletter focused on practical AI tutorials for people who’d rather replace than be replaced by AI.

Part 1: For all Claude users

#1 Separate the work from the system

When working with AI, you usually have two kinds of folders.

The first kind is the work. The docs, the images, the videos, the spreadsheet you track them in. Things a person opens, reads and uses. I call this content, and it is what I share with my team.

The second kind is the system: the skills, the scheduled tasks, the prompts. Nobody reads those for pleasure (but Claude). They are the instructions that produce the work, and they behave far more like code than like content — edited constantly, broken by one wrong character, and worth nothing unless you can see exactly what changed and go back.

When I started using Claude, I kept both in the same space for months. It was the first thing I fixed, because the two need opposite things: content wants to be easy for people to browse and reorganise, and the system wants to be left alone and versioned.

So now I have two workspaces:

• Content workspace: the files Claude generates that I share with my team

• System workspace: sits on my own machine with version history (more about this in lesson 9)

I want to start easy, so everything in this lesson is about the content side, because that is the side you and your team actually open.

Here are the good practices I follow to organize the content workspace:

1. Start by numbering every folder in it using the format NN-name

01-clients
02-invoices
03-archive

Why? The order stops being alphabetical and starts being the order you actually work in.

The number can carry more than order. Low numbers for what I touch daily, higher ones for what I open occasionally:

01-clients          I open this daily
02-invoices         I open this weekly
03-archive          I open this a few times a month

Now the folder your team shares holds only the things your team opens.

2. Keep folders you might never open out using 99-<name>

My workspace/
  01-content/          what the team opens: clients, invoices, archive
  99-vault/            records kept. I almost never open this

Same numbering trick, one level up. The split means someone reorganising the content folder physically cannot reach those files.

The same trick names the steps of a job:

01-drafts
02-review
03-done

Now the position of a file is its status. Moving a file into the next folder becomes the act of approving it.

Skills, scheduled tasks, prompts are closer to code than to content. That’s why they live in one folder outside of this workspace, with version history, and we’ll see later why that matters more than it sounds.

Subscribe now

#2 Name folders, skills and tasks for what they hold

I had a folder called claude. Everything AI touched lived in it. It made sense at the time and it was wrong, because the name described the tool instead of the contents.

Tools change. You might use Claude today and something else next year, or both at once. The moment you do, the folder called claude is full of things that have nothing to do with Claude, and you either rename it or live with a lie.

Name things for what they hold.

Bad:   claude/          chatgpt-stuff/       ai-output/
Good:  system/          clients/             drafts/

In the case of jobs (skills, scheduled tasks, etc), the names should say what they do: what goes in and/or what comes out, or the action itself if there's no clean input/output. The ones that rot describe when it runs or what version number it happened to be.

Rots fast:   step-2-processing/      monday-batch/       new-system-v2/
Ages well:   invoice-check/          leads-to-crm/       backup-database/

"Monday batch" stops being true the first time you move it to Tuesday. "Step 2" stops being true the first time you add a step.

I learned this from a task named after two kinds of work it handled. One was retired within a month, and the name outlived the thing it described, which is the worst outcome, a name nobody trusts and nobody dares change.

Two things follow from this.

1. Use the same name everywhere: In Claude, a single job usually has three names: the skill that holds the instructions, the scheduled task that runs it, and the label you see in the app. I used to have different names for the same job. Every time I opened the app, I translated between three names for one job. Now I give the skill, the task, and the label the exact same name. If I ever need to find something, I only have to remember it once.

2. Rename when the name is actively misleading, not when it is merely not your favourite: Misleading is maintenance. When you do rename, rename everything at once (so nothing breaks).

#3 If you’ll edit it later, use variables

This one is about scheduled tasks. They run on their own, at a time you pick, without you there.

They’re great, but have one problem. Cowork isn’t a flawless product.

Anthropic ships updates almost every week, which is good, but that also makes Cowork unstable. One day, Claude updates a scheduled task with no issue. The next day, it tells you it doesn’t have permission to touch the task. Then you open another chat and it works again.

Now imagine your task is one long paragraph with every value (the folder, the time, the account, etc) written inside the sentences. Every time you want to change one, you are editing prose while fighting the app.

Here is what that looks like:

<more instructions>
-----
Schedule my posts from the "done" folder to artificialcorner.com at 5:00 and 21:00,
look 4 days back, and use the Chrome profile <xyz>.
-----
<more instructions>

Every value that task needs is buried in that sentence, somewhere in dozens of lines.

The fix is to move every value that could change in the future into a block at the top:

SOURCE_FOLDER   = 03-done
POSTING_TIMES   = 05:00, 21:00
DAYS_BACK       = 4

Everything below refers to those names instead of repeating the values. Now changing where something lands is a one-line edit at the top. No new chat, no permissions dance, nothing rewritten.

Don’t do this by hand.

Run the scheduled task you want to fix, stop it and paste this prompt:

Turn the hardcoded values in this scheduled task into variables at the top. Show me the best candidates and wait for me to pick which ones actually go in.

Claude might suggest more candidates than you need. Pick only the ones that you’ll actually change over time. The rest stay in the prompt.

Note: Do not turn the whole thing into placeholders. This is the opposite mistake, and I made it straight after learning the first one. Once every detail is a variable name, the instructions read like an empty form, and the results might get worse because the model has nothing concrete to anchor on.

#4 If you have many scheduled tasks, build a calendar

Once scheduled tasks start working, you add another one. Then another.

I have 20+ running right now.

And that’s where the Scheduled screen stops helping you. It shows cards sorted by next run, and each card gives you one line: “Weekdays at ~6:00 AM.” That’s enough at three tasks. At twenty, you’re guessing.

But losing track is the small problem. Here’s the expensive one: usage

Claude doesn’t only limit you by week. It limits you in rolling 5-hour windows. So if six tasks fire between 6:00 and 8:00 AM, they don’t just run. They land in the same window and eat it together. Then you sit down at 9, ready to work, and your window is already spent (on tasks that could have run at any time).

Here’s how to see your real schedule (so your usage limit lasts longer).

Open Cowork and paste this:

Read all my scheduled tasks and build me a one-day calendar as an inline visualization.

Layout: 24 hourly slots in a grid, 4 per row. Skip paused tasks. Tag weekly tasks with the days they run (Mon, Thu) and monthly ones with the day of the month (day 1, days 13-18). Mark empty hours as empty.

Color-code by category.

Here’s part of my calendar created by Cowork:

Most of my scheduled tasks run while I sleep. I stack the small ones at 6 AM on purpose, because they barely touch the limit.

If you’re on Claude Code, you don’t need to do this. Just go to Routines → Calendar.

#5 Make related jobs report the same way

I have dozens of jobs running overnight and fifteen different ways of telling me what happened. Every morning I used to read fifteen shapes and hunted for the one line that mattered in a different place each time.

Now they all end the same way:

✅ DONE — <task>          
​2 of 3 reports built

Made
| Item | Where | When |
| Q3 summary | /reports/q3 | Fri 08:00 |
| Cost breakdown | /reports/costs | Fri 08:04 |
​
Held back
  ...what did not go out, and why
​
⚠️ Needs you
  ...the one thing only I can do

Three states at the top, and only three: done, failed, nothing to do. Then what it made, in real names. Then what needs me, always last (so I can always quickly find it at the same place)

A section only appears when it has something in it. A line that shows up every single day has stopped carrying information, so I let its absence say it instead.

Here the prompt that sets this up:

Create a skill called report-format that defines how my scheduled jobs end

their run. It should say:

- Open with two lines: a state (done, failed, nothing to do) and the job

name on the first, one short clause saying the outcome on the second.

The state emoji is the first character of the report.

- Then “Made”: a table, one row per thing produced, with real names, links

or times. What it made, never how it made it. When one item went to

several places, give that item its own small table with its name in bold

above it, rather than one big table with a repeated name column.

- Then “Held back”: anything that could have gone out and did not. Two

columns, Item and Why.

- Then “Needs you”, always last: one specific action I can take in one reply.

- A section appears only when there is something in it. A clean run has no

heading at all, and never the word “nothing”: its absence is the answer.

I read these on my phone, in a column about forty characters wide. That is

the only width this format has to survive, and two rules follow from it that

outrank everything else in the skill:

- Never wrap the report in a code block. A fenced block renders monospace

and does not reflow, so every line wraps in the middle of a fact. Fences

are for a literal command or a path, never for the report itself.

- Never align anything with spaces. Columns built out of padding hold only

at full terminal width. Structure comes from markdown, which reflows.

Keep the tables inside that width: at most three columns, headers of one or

two short words, short cells. Trim the type prefix off a folder name. Dates

as “Sat Sep 12 21:00”, counts as “5/12”. A long quoted title never goes in a

cell - put it under the table or leave it out. If an item needs a fourth

fact, that fact goes in a sentence under the table, not in another column.

Not everything is a table. The explanation on a failed run is prose. So is

anything the table cannot carry. And a “smells” section, if you add one, is

bullets - a smell is a sentence, and a sentence in a cell wraps badly.

After you create the skill, add this section to the target jobs:

## Reporting

Report with the “report-format” skill.

Note what it is doing: the format goes in one place, and each job points at it. Do not paste the format into every job, or you’ll be editing fifteen files every time you change your mind.

The shape is for jobs that belong together. A one-off task you run by hand has nothing to be consistent with, and forcing it into the same template just adds ceremony.

A note on smells: Jobs notice things that are worth knowing but are not actions. Give them their own section, and give it a threshold in the same sentence you ask for it: "Flag it only if it could make this report wrong, or it will change a future night. Never flag a decision I made on purpose." Keep smells out of the verdict, too. A job can flag them and still have done its work perfectly.

#6 Tell Claude to delete what it removed

This one is entirely a Claude habit, and it took me a while to see it.

When Claude removes something from a skill or task, it likes to leave a tombstone: “option 3 — removed”, “renamed from monday-batch”, “we no longer split these two”. It is being helpful, in the way a changelog is helpful.

In an instruction file, it is not as helpful, though.

Claude reads the whole file as instructions. A line saying “we no longer use option 3” still puts option 3 in front of it. Mine reached for a deleted option twice in one month — the option only existed as a note explaining that it did not exist.

To fix this, use this prompt when you have Claude Cowork edit a skill or task:

When you remove something from an instruction file, delete it completely.
Never leave a note saying what it used to be, what it was renamed from, or
that it was retired. 

The file states only what is true now.

In Claude Code, that line goes in claude.md.

That stops new ones. For the ones already in your files, do a one-time sweep:

Read my instruction files and list every line that describes what something
used to be or no longer does. For each one, show me the line and the file, and tell me whether anything still references it.

Do not delete anything yet.

After that, drop anything that is no longer useful.

Click here to get my FREE Claude course (20+ lessons)

Subscribe now

Part 2: For Claude Code users

#7 Every line in your CLAUDE.md gets billed forever

CLAUDE.md is the file Claude reads before it reads anything else.

But most people stop there. They never look at how the file should be written.

You can type the following command to create claude.md inside your Claude project.

/init

I did it.

The claude.md is ready.

But if your claude.md is too large, remember that every conversation with Claude sends those instructions to your agent.

For instance, I don’t think this information needs to be sent at the start of every conversation. But first, I measured its impact with this prompt:

Calculate the tokens this CLAUDE.md needs and tell me in one sentence.

Here it is.

Now, let’s make it thinner using this prompt;

Audit my CLAUDE.md. Keep only the lines that are true in every session.

Move task-specific lines into skills, and delete anything outdated, duplicated, or already enforced by a linter or hook.

Rewrite what survives as commands and hard rules instead of prose, then tell me the line count before and after.

After Claude Code rewrites it, look at how many tokens the file needs now.

More than 50% save, using one prompt.

The rule: CLAUDE.md holds what is always true. Everything conditional goes in a skill, regularly update yours to save tokens and time, so lower the cost and use the limit more.

#8 A rule in CLAUDE.md is a suggestion — a hook is a guarantee

This is the feature I discovered most recently. The rules written inside claude.md can be followed, but when a session gets long and the context fills up, Claude may suddenly stop following some of them.

Hooks are different. If you set one, Claude can enforce that rule much more consistently.

These days, writing with AI is almost treated like a sin across the web. Substack now gives readers the option to scan text with AI, while LinkedIn has added an AI slop button. YouTube and X are also taking action against AI slop.

So if you write with AI, you need to be careful.

One of the best sets of rules for avoiding AI slop was written by and shared in this GitHub repo: https://github.com/petergyang/no-ai-slop

Let’s turn these rules into a hook, so every markdown file Claude writes gets checked against them.

Paste this prompt:

Set an instructions for AI sloop using the info from this github: https://github.com/petergyang/no-ai-slop

Now the skill is ready.

It is time to add this as a hook. To set this hook, use this prompt.

Create a PostToolUse hook that checks every markdown file I write against my writing rules.

The rules live in no-ai-slop skill

Read the banned words and banned structures from there, and block any note that breaks them.

Test it by writing a note that breaks two rules on purpose. Show me the catch and the fix.

It’ll ask your approval to update your files, and approve it.

Claude already set up the hook for us.

Now, before Claude’s answer reaches us, check-slop.js checks it and removes the AI slop.

If you put the same rules inside CLAUDE.md, some of them may be ignored as the context gets longer. Hooks are much harder to bypass.

Subscribe now

#9 Learn just enough git

I put off learning git for years (I regret that). Then I reorganized my whole system in a week and git was the safety net that made it possible.

What it is in plain English: Git takes a “photo” of a folder whenever you ask. Every photo is kept, forever, with a note from you about what changed and why. You can compare any two photos, and you can go back to any of them.

That’s it. Everything else is detail.

Why it matters when you work with AI: You are letting Claude edit dozens of files at once. Without git, “undo” means remembering what it touched. With git, “undo” is one command and it is exact.

The four things worth learning:

git status     What have I changed since the last photo?
git diff       Show me those changes, line by line.
git add -A     Include everything.
git commit     Take the photo, with a note about why.

Nowadays, you can make Claude run those commands for you using plain English prompts. That said, it’s worth knowing what’s going on behind the curtains.

#10 Delete half your MCP servers

MCPs were great when they arrived, but installing one meant editing a JSON file, which was unusual for most of us.

Now, you can connect them to Claude much more easily using Connectors.

All you have to do is pick one from the list above, or you can add a custom one too.

As you can see, the setup takes more than one step. And because of how MCP works, when you connect an MCP server, its tool definitions are added to the context even if you do not use them.

On my own setup that came to 181.1k tokens across 298 tools, none of them in use. I show the screen in tip #7.

Most of these services also have a CLI version. It often does the same job, can be installed with one prompt plus authentication, and the CLI approach existed long before MCP or AI.

My favorite CLI is NotebookLM, and you can install it with just this prompt:

Install this CLI from this link: https://github.com/teng-lin/notebooklm-py and let me know when I have to authenticate, run the authentication code for me and open the browser so I can.

After you authenticate, you can control NotebookLM directly from Claude Code.

The rule: If a tool has a CLI, use the CLI. Keep MCP for services that have no other option.

#11 Your best model should judge, not type.

You probably run your prompts using the latest model, unless you hit your Fable 5 limit.

If you are a regular Claude user, you know the feeling.

But for most of the tasks, you don’t need the strongest model. To do that, we’ll use subagents.

For instance, Sonnet 5 drafts well enough, and it does not eat your weekly limit the way Fable does.

So hand the drafting to three of them and keep Fable 5 on top as the referee.

Here is the prompt template you can use.

Create three subagents in .claude/agents/ with model: sonnet, each taking a different angle on [TASK].

Give them the same brief and run them in parallel.

You stay on Fable 5 and you never draft.

Read the three drafts, judge them against [CRITERIA], pick one, and say what is wrong with it.

Send it back for a revision instead of fixing it yourself.

Customize the task and criteria, and your limits will work much better.

The rule: cheap models produce, the expensive model reviews.

Never spend your strongest model on work; a weaker one can redo for free.

#12 You are running one Claude. Run twelve.

Most people open Claude Code, type a task, and watch one agent work through it in order.

Meanwhile, they switch tabs and lose focus, or worse, one agent carries a single set of instructions for every step of the job.

That is not how the people who build these tools work. Here is one example that runs more than one Claude.

Let’s say you are building a landing page for your newsletter. The parts you see are called the frontend. When you click a button or submit a form, the backend starts working and runs the required functions.

So, when building one, you may want two agents: one for the frontend and one for the backend. Let’s paste the following prompt to do this:

Create two project subagents in .claude/agents/, then use them to build a landing page for LearnAIWithMe, a newsletter signup.

First read https://learnwithmeai.com/ and pull the positioning, the tone,and what the newsletter actually offers. That is the source for the copy.

Do not invent claims that are not on the site.

Then write the agent files yourself. Give each one a name, a description, a tool list, and a system prompt. Split them by file ownership so they never touch the same file:

- frontend owns site/index.html and site/style.css, and needs WebFetch

- backend owns site/server.js and site/api/

In each system prompt, write the boundary as a rule the agent cannot cross,and tell it to return a one-line summary when it finishes. The contract between them is POST /api/subscribe, which takes an email field.

Then run both in parallel. Frontend builds the page from the real copy. Backend builds the endpoint and stores the emails. Do not write any of this yourself in the main session.

Just after you paste this prompt, Claude Code will ask your permission, but the most important part is writing the agent files; approve this one.

Next, it’ll ask your permission to write instructions for these two agents.

This means you’ll have one specialized agent for each of your website parts, frontend and backend.

And they run in parallel, so the backend is not waiting on the frontend to finish.

The website is ready.

The rule: If your task can be done by creating multiple subagents, just tell it to Claude. Don’t do everything with the main agent.

Subscribe now

Bonus: Four commands nobody uses.

Here are 4 commands worth learning.

/context shows what is sitting in your window right now, broken down. Let me show you.

/usage shows you the usage limits.

/save-memory writes the correction into your memory file without leaving the session. So when Claude does something wrong, you fix it in the same breath. Later never comes.

/compact will summarize your earlier conversation inside the project.

That’s it! If you have a lesson to share, leave it in the comments

Share

Join my newsletter with 80K+ people to master AI in 2026

Let’s connect! You can find me on LinkedIn, Instagram and X

Click here to get my FREE Claude course (20+ lessons)

This is part 2 of a three-part Claude Cowork series. Don’t miss part 1 and part 3.

Applying to jobs is a numbers game, and that’s exactly the problem.

You open a post, read it, decide if you’re a fit, tweak your CV, write a proposal, send it. Then you do it again. Do that fifteen times and your day is gone.

Right now I have 161 job posts sitting in one dashboard.

161 jobs you could apply to. And in a few days, it’ll be hundreds.

That’s good news until you realize someone has to read them all. So why should that someone be you?

A Claude Skill is built for exactly this kind of work.

So we built one and named it Head Hunter. It reads your CV, scores every job post against it, and writes proposals for the 10 best ones each day.

Before we run it, let's cover the fundamentals:

• What a skill is, in plain English

• How to create skills & how to call them

• The Head Hunter Skill (skip to this section if you know the basics)

This is the second part of my Claude mini-course: 3 guides to automate your job, your life, and pretty much everything else with Claude Cowork (read part 1 here)

Let’s start with part 2: Claude skills

I made this mini-course in collaboration with . He 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

Subscribe now

What a skill is, in plain English

A Claude Skill is a Markdown file that includes a set of instructions.

Once you start using them, you don’t have to save prompts all the time. And there are plenty of free skills available on the web.

Here is one website called skills.sh, with 1.3 million skills.

And the #1 most downloaded skill is “find-skills,” which helps you find the right skills for your task. But how to create one?

How to create skills?

You can create a skill by uploading an .md file, creating one by prompting, or using the /skill-creator command.

All three work.

Let me show you each of them quickly.

#1 Creating a skill by prompting

Let’s say you pick one skill from skills.sh and want to install it.

Here is the fastest way.

Install this skill from this link: [paste-link]

Let’s install Matt Pocock’s “Grill Me” skill using this method and the following prompt.

Install this skill from this link: https://www.skills.sh/mattpocock/skills/grill-me

In 5 minutes, Claude installed it for us.

All you have to do is restart the session and start using it.

#2 Creating a skill using /skill-creator

This is a slash command for creating a skill, but it can also modify and improve an existing one.

It can run evals to test the skill, find gaps, and even benchmark it.

To use it, all you have to do is explain what you want the skill to do, as I did with this prompt:

/skill-creator: I want to create a skill hunts job

After pasting the prompt, it’ll ask you a couple of questions based on the task you want it to handle, then create a skill that fits your needs.

Skill Creator found my Head Hunter skill, read it, and started asking questions about what it saw.

In the end, it asks me to run an eval loop, which runs the skill against test cases and reports where it slipped.

I already created a Head Hunter skill, so I won’t go through that step again.

But after creating a skill, you need to close the app and reopen it before you can run the skill.

Subscribe now

#3 Uploading an MD File

Let’s say you want to install the Head Hunter skill. Uploading it is one of the easiest ways.

Click here to download the Head Hunter Skill

To do that, first, go to Settings > Capabilities > Skills > Customize inside your Claude App.

Next, click on “Add” and “Upload Skill”.

And drop the .md file here.

And you’ll see the new skill in the list.

Click here to get my FREE Claude course (20+ lessons)

How to call skills?

The fastest way is using the slash commands.

For instance, when I type /head-hunter, it launches the skill I built to hunt jobs for me.

Let’s do this.

Head Hunter Skill

Quick context before we run it. The 161 job posts come from Part 1 of this series, where Claude in Chrome scraped Indeed, Glassdoor, and Upwork into one Google Sheet.

Head Hunter reads your CV first. Then it scores every job in the sheet against it, writes a proposal for the ten highest, and drops all ten into one Google Doc.

Click here to download the Head Hunter Skill

How does it work?

Every job gets a score out of 100, built from four parts.

• Skills overlap, 40 points.

• Seniority fit, 25.

• Domain fit, 20.

• Practical fit, 15, which covers location, remote policy, and salary.

Two rules do the heavy lifting.

A tool you listed from a course scores zero; only production evidence counts. And when a post hides the salary or the seniority, that part scores low instead of getting a guess.

This is the part worth editing. Disagree with a score, open the file, change one weight, and every future run follows the new rule.

So you call it with a slash command and leave it alone.

In 20 minutes, it analyzed all 161 jobs, scored them, matched them with my CV, and started creating a Google Doc with proposals for the 10 best-fit jobs based on those scores.

The folder already contains the scraped jobs and my CV, so I don’t have to add any extra explanation.

And finished.

Let’s check this folder.

I clicked the first one.

Subscribe now

Next Steps

Ten proposals, sitting in a Doc, from a folder I never had to explain.

Now, imagine scraping these jobs every day and generating proposals for each one. All you need to do is polish the proposals and send them.

How many hours would that save someone who is actively job hunting?

Well, we’ll do exactly that in Part 3, using Claude’s scheduled tasks.

See you in the next one.

Join my newsletter with 80K+ people to master AI in 2026

Let’s connect! You can find me on LinkedIn, Instagram and X

When Claude hands back something you have to rewrite, there’s an obvious suspect: the prompt.

So you rewrite it. You add detail. You try phrasing it the way you saw someone phrase it online.

That works for a while. Then it stops working.

What finally moved things for me was paying attention to the opposite case: the sessions where Claude got it right on the first try.

The pattern was the same every time.

Claude had context it could reuse. The files were organized. It understood the task before starting. And the complex work was split into stages instead of one giant request.

Not one of those is a prompt trick.

They’re all good habits I’ve been following over the past few months:

Subscribe now

Read more

Click here to get my FREE Claude course (20+ lessons)

This is part 1 of a three-part Claude Cowork series. Don’t miss part 2 and part 3.

Hi, it’s Frank. Hope you’re enjoying your summer!

This month I’m releasing a Claude mini-course: 3 guides to automate your job, your life, and pretty much everything else with Claude Cowork.

I’ve built dozens of AI automations by now. Only 3 features show up in at least 80% of them. We’ll cover all 3 in detail across these guides.

This mini-course is about getting things done with Cowork, without the fuss.

No fancy setups.
No complex systems.
Just the essentials you need to begin.

That’s literally how I worked during my first months with Cowork. My workspace was messy (still kind of is). It didn’t follow any best practices. But it got things done.

Once you finish this mini-course, you can move on to my other guides and pick up the best practices from there.

I made this mini-course in collaboration with . He 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

The use case we picked for this series is job hunting. That said, the core concepts apply to any automation you build in Chrome.

Let’s start with part 1: Claude in Chrome.

Every job hunter has the same morning routine.

You wake up, make your coffee, and open job search websites like Glassdoor and Indeed, then Upwork for freelance work.

You scroll through the same listings you saw yesterday, while a new posting is hiding. And if you are searching across multiple websites, it gets even worse.

You might see the same job on all three platforms. Sometimes, you see it twice on the same platform. Or you may find the same role posted by both the company and a recruitment agency.

Job searching already requires the right mindset, and trust me, this can get frustrating very quickly.

The solution? Automate it.

We can use Claude in Chrome to automate anything you'd normally do in a browser. Here's how.

Subscribe now

How to set up Claude in Chrome

Claude in Chrome is a browser extension. It puts Claude inside the browser you already use. It looks at the page the way you do, then scrolls, clicks, and types where the page needs it.

Before you start, make sure you have the Claude desktop app installed and that you’re on Cowork (not chat).

How to activate Claude in Chrome?

First, go to Settings, then Claude in Chrome, and click Install. When it finishes, Chrome asks you to add the Claude extension.

In the permissions, there are two options: one is allowing all sites or blocking them.

You can allow all websites and block the ones you want to block, like this.

Don’t forget to install the Claude in Chrome extension in the browser: here’s the link.

Click here to get my FREE Claude course (20+ lessons)

Indeed goes first

Now we can test by scraping job listings from Indeed.

Here is the prompt to paste.

Use my open Indeed tab. Search for [ROLE] jobs, location [LOCATION / remote],

posted in the last 7 days.

Scroll through the results. Click into every posting that matches, and collect: title, company, date posted, salary if shown, and the posting URL.

Some jobs appear twice, once from the company and once from an agency.

Keep both, but flag the duplicates you notice.

Do not apply to anything. Do not change my account settings.

When you are done, return one table with every posting, newest first.

You can customize the role, location, and whether the job is remote or on-site. After pasting the prompt to scrape jobs from Indeed, it asks for my approval.

Now Claude is using the browser.

Here is the part of it.

After 13 minutes, let me show you the report.

As you can see, it scrapes job posts from the last 7 days and even saves them to a Google Sheet in my Google Drive.

This is only for Indeed.

Subscribe now

Now Glassdoor and Upwork (same sheet)

Now, let’s do the same with Glassdoor and Upwork, using this prompt.

Use Claude in Chrome. Two sites this time, one after the other.

First, go to my open Glassdoor tab. Search for [ROLE] jobs, location [LOCATION / remote], posted in the last 7 days. If Glassdoor asks me to sign in, pause and let me log in, then continue.

Then switch to my open Upwork tab. Search for [ROLE] projects posted in the last 7 days. Upwork loads more results as you scroll, so keep scrolling until nothing new appears.

From both sites, collect the same fields as before: title, company or client, date posted, salary or budget if shown, and the URL. Add a column called Source with the site name.

Append everything to the Google Sheet from the Indeed run. Do not create a new file. Do not overwrite the Indeed rows. Some of these jobs already exist in the sheet from Indeed. Keep them, but mark the duplicates you notice in a column called Duplicate.

Do not apply to anything. Do not submit any proposal. Do not change my account settings.

When you are done, tell me how many postings came from each site and how many duplicates you flagged.

Claude opens Glassdoor.

And it starts to scrape.

Next in line, Upwork.

But it looks like it hit the wall while scraping through Upwork.

I steered with this prompt

For Upwork, you can save it as HTML and scrape it if you hit a wall.

But it appears it already solves this problem.

It scrapes 161 total jobs.

Turning the data into one dashboard

Let’s turn the results into a Dashboard so we can see them in a cleaner, more visual format, using this prompt.

Read the Google Sheet from the job runs. Build a single HTML file called job-dashboard.html that shows every posting in one view.

At the top, four numbers: total postings, postings per site, how many are new in the last 48 hours, and how many duplicates.

Below that, one table with every posting. I want to sort it by date, salary, or source by clicking the column headers, and filter it with a search box. Duplicates stay visible but marked, do not delete them.

Salary is the messy column. Indeed writes yearly ranges, Upwork writes hourly rates or fixed budgets. Do not convert them into one number, show them as written, but add a Type column that says salaried or freelance so I can filter on it.

Use the exact data from the sheet. Do not invent, merge, or clean any row beyond what I asked. One file, no external libraries, works offline when I double-click it.

And here is the dashboard.

Subscribe now

What’s next

The search now runs from a single prompt, and every posting sits in one dashboard. In the next guide, we’ll level up this automation using Claude Skills.

Join my newsletter with 80K+ people to master AI in 2026

Let’s connect! You can find me on LinkedIn, Instagram and X

Click here to get my FREE Claude course (20+ lessons)

Claude just released a new feature called “Record a Skill.”

A Claude skill is a set of instructions you save once. Claude then reuses them every time you need that same task done. Think of it as teaching Claude a job, then never explaining that job again.

I've created dozens of skills the old way. You'd type out every single step to teach Claude how to handle a task. That's tricky if you're not used to describing a process in writing. It's slow if you'd rather not type it all out.

Record a Skill is the easiest and fastest way to create a Claude skill.

With this new feature, you just record yourself doing a task and let Claude build the skill from what it observes.

In this guide, you’ll learn:

• How to use Record a Skill

• How to improve any skill

• When to use it (and when the old way still wins)

Subscribe now

First things first: Find the Right Task

Record a Skill works best for tasks you do with your hands more than your head. Anything click-heavy, repetitive, or full of small habits you'd forget to mention if you typed it out.

If you'd rather show someone than explain it, this feature is for you.

I have a few of those myself. Tasks I'd rather show than type out, which is exactly why some sat untouched for months. One of them sat on my list since March and I’ll use it as an example for this guide.

First, here’s some context:

A couple of times per week, I record myself doing some gym exercises. The videos sit on my iPhone, and when I move them to my Mac they all land with generic names (IMG_4521, IMG_4522, and so on). That’s useless to me.

Without Record a Skill, here’s what I’d have to type to create a Claude Skill:

This folder contains videos of me doing gym exercises.

The videos contain generic names. I want every file name to tell me three things: 

- the exercise
- the weight I used (if visible on the video)
- the number of reps

...

[more instructions]

...

Once the files have real names, I want them grouped into folders by muscle group.

Pull-ups go in the back folder.
Squats go in the legs folder.
Tricep exercises go in the triceps folder

...

[more instructions]

This is just a simplified version. Notice I only said what I wanted, not how to get there. Sometimes Claude figures out the how on its own. Other times, you have to spell out every step.

I postponed this because it was not important enough to justify the typing. Once Record a Skill was released, I quickly created the skill by showing Claude the steps I’ve been doing manually for months.

Subscribe now

Record a Skill

Follow these steps to use this feature:

Step 1: Open the Claude desktop app → Home → Cowork → + → Record a skill

Step 2: Show Claude how you’d do the job

In the video below, I show Claude the steps I follow to rename my video files: I watch the video, identify the exercise and weight, and count the number of repetitions. This is easy to do when there’s one file, but time-consuming when you’ve recorded dozens of videos.

Keep this in mind for step 2:

• Before you record, close any files, apps, or conversations you don’t want captured.

• A recording can run for about 10 minutes. A countdown appears in the capture bar when you have about a minute left

• Make sure your recording is clean. Any mistakes or false starts may confuse Claude

• If necessary, talk through what you’re doing while you record. Narration gives Claude context it can’t get from your screen alone.

Step 3: Stop the recording and let Claude generate the skill

Once you’re done, stop the recording.

Claude will automatically generate the task, review the recording, then propose a skill.

After a couple of minutes, a “save skill” box will pop up. Click Save.

You’ve just created a new skill!

Treat this skill as version 1. If necessary, improve the skill following the instructions below.

Subscribe now

How to improve any skill

Record a Skill gives you a nice first version.

That might not be enough, though.

After a few runs, you might find things the recording didn’t capture: exceptions you forgot to show, rules that only exist in your head, and steps where Claude guessed instead of knowing.

Here’s what happened with my gym skill. The first version renamed the videos correctly, but it got confused with two edge cases:

• Exercises with a bar (the bar weighs 20kg, but the number never appears in the video)

• Bodyweight exercises like pull-ups (there’s no weight at all)

The fix is simple. The improvement process for any skill has 3 steps.

Step 1: Run the skill

Once the skill is created, open a new task and run it. Type / followed by the name of the skill to call it.

The improvement happens by running the skill.

Don’t expect the first run to be perfect. Expect some back and forth:

• Claude realizes things on its own when working with new samples

• You give Claude feedback (”bodyweight exercises should say Body Weight instead of a number”)

• You correct, Claude retries, you correct again

All of that back and forth is not wasted time. It’s the raw material for the next two steps.

Step 2: Extract the lessons and your feedback

Once the run finally ends successfully (and only then), I send these two prompts:

List the hard-earned lessons from this run.

Claude goes through everything that happened during the task: what it tried, what failed, and what it had to figure out on its own.

How about the feedback I gave you? List them too.

This catches all the corrections you typed mid-run. Without this prompt, those corrections die with the conversation. Next run, Claude makes the same mistakes again.

Why extract these lists instead of jumping straight to updating the skill? Because it lets you review them before they go in. Sometimes Claude learns the wrong lesson from a run, and this is your chance to catch it.

Step 3: Fold everything into the skill

Once you’re happy with both lists, send the final prompt:

Update the skill: fold my feedback and the hard-earned lessons into the skill.

Claude edits the skill file for you. Next run, those cases won't trip Claude up anymore

Click here to get my FREE Claude course (20+ lessons)

When to use it (and when not to)

Record a Skill isn’t replacing the traditional way of creating skills.

The traditional way: you open a task, tell Claude what you want, go back and forth until Claude handles it exactly how you like, and then ask it to turn everything into a skill.

Both roads end at the same place (a skill). The difference is where the knowledge lives.

Record it when the knowledge lives in your hands

Some tasks you do on autopilot but would struggle to write down. Try describing how you tie your shoes. Hard, right? Doing it takes 3 seconds. These tasks are perfect for recording:

• Click-heavy tasks in apps and menus, especially when they jump between multiple apps

• Tasks full of micro-decisions you don't even notice you make: duplicating a file before touching it, pasting as values instead of regular paste, etc. You might forget to type these, but the recording (plus your narration) catches them

• Tasks you’ve been postponing for months because typing the steps felt like homework

Type it when the knowledge lives in your head

Some tasks have nothing to show on screen or are too complex. For these, the chat route wins:

• Complex tasks. Some of my most useful skills came from long chats. These aren’t click sequences. Some are decision trees full of rules and checkpoints. Some are problems where clicks were never the answer (Claude created a small script to solve them). You can't record any of that. You can only talk your way there.

• Workflows that don’t exist yet. You can’t record something you’ve never done. In a chat, you and Claude design the process together

• Tasks with many “it depends” moments. A recording captures one successful path. In a chat, the back and forth surfaces the branches before the skill exists

• Tasks where your manual way is a workaround. If you copy-paste between 4 tabs because that’s the only way a human can do it, don’t teach Claude your detour. In a chat, Claude often proposes a faster route (a quick script instead of 50 clicks)

My rule of thumb: imagine you have to explain a task to a coworker over the phone. If you’d say “it’s easier if I just show you,” use Record a Skill.

One more thing

You’ve read the whole process. Recording, running, fixing, running again.

Don’t let it stay theory.

Go automate that task you’ve been postponing since who knows when.

I wrote this guide so you can teach Claude the way you'd teach a new hire: by showing, not writing a manual. Becoming a paid subscriber gets you my full library of guides, the skills and workflows I actually use, and new ones every week.

Join my newsletter with 80K+ people to master AI in 2026​

Let’s connect! You can find me on LinkedIn, Instagram and X

Subscribe now

Hi!

Today, is sharing with us a guide to NotebookLM (my favorite AI tool for learning new subjects)

Jeremy is a tech curator and educator who writes Wonder Tools, a weekly Substack newsletter that helps you discover the most useful sites and apps for creative productivity. He also serves as Director of Teaching and Learning at CUNY’s Newmark Graduate School of Journalism in New York City.

Click here to get my FREE Claude course (20+ lessons)

NotebookLM is the most useful free AI tool I've tried. It has twin superpowers. You can use it to find, analyze, and search through a collection of documents, notes, links, or files. You can then use NotebookLM to visualize your material as a slide deck, infographic, report — even an audio or video summary.

This week, Google renamed NotebookLM to Gemini Notebook. Same product, but it’ll do more across the Google ecosystem.

Subscribe now

Getting Started with NotebookLM

Go to NotebookLM → Try NotebookLM → Create new notebook

At first sight, it looks similar to ChatGPT or Gemini

1. In “sources” you can see your uploaded materials (PDFs, docs, etc)

2. In “chat” you can chat with it (just like any other chatbot)

3. In “studio“ you can transform uploaded sources into different content formats

NotebookLM is more powerful than it looks.

The fact that its responses are grounded in the materials you upload makes it less likely to hallucinate about those topics than ChatGPT.

How to set up a notebook

1. Pick a purpose. Start a new notebook for a work project or a learning goal. Examples: I created a notebook to organize materials for the new online bilingual MA program we’re developing at the CUNY Newmark Grad School of Journalism where I work. I also set up a notebook to learn more about Gustav Mahler, a composer I revere. I have numerous others for work and personal projects.

2. Find sources for your notebook. NotebookLM recently added a search panel to help you discover high-quality sources. You decide which, if any, of the suggested materials to add to your notebook. The “Fast Research” is quick and focused, unlike a generic Google search that returns hundreds of results, some of which have gamed the search engine system.

• Fast Research surfaces 10 or so documents related to your topic in less than 30 seconds. You can ask it to find sources within your Google Drive, or from the Web.

• The Deep Research prompt option in the same panel will more slowly gather many more sources.

Tip: make your query as specific as possible to surface relevant, useful sources. Here’s an example of a concise, precise query I used.

3. Add your own materials. Upload files up to 200 MB and 500,000 words into your notebook. You can add:

• Google Docs, Slides, and Sheets

• PDFs, images (including photos of your handwritten notes), and Microsoft Word documents

• YouTube links and audio, image, or video files (it extracts the transcript)

• Website URLs (it extracts the text)

No other AI tool I’ve used lets you compile as many different kinds of materials in a centralized AI workspace that’s easy to explore and build with.

• Free accounts can create up to 100 notebooks, with 50 sources in each. On a free plan, you may run into limits when creating multimedia materials. You can run free 10 Deep Research queries a month. Students in the U.S. 18 or older can get pro access for free.

• Pro accounts, which cost $20/month as part of Google AI Pro, can host 500 notebooks with 300 sources in each. They can run 20 Deep Research queries a day.

Subscribe now

Collaborate and share

NotebookLM now lets you collaborate as you would with Google Docs. You can choose to invite people as viewers or editors. Give them a full view of your sources and notes, or limit their access to the search/chat interface.

You can also publish notebooks publicly. Here are some examples:

• Trends in health, wealth and happiness by Our World in Data

• How to build a life, from The Atlantic

• Shakespeare’s Complete Plays

• Parenting Advice for the Digital Age, by Jacqueline Nesi, PhD of Techno Sapiens

• Earnings Reports for the World’s 50 Biggest Companies

• Secrets of the Super Agers by Eric Topol

Explore your materials

As you add materials, NotebookLM analyzes them and suggests relevant questions. After I uploaded biographical material about Mahler, it suggested search queries — based on the source documents — about why he converted to Catholicism and what poetry collections inspired him. You can also ask any question on your mind or type in any kind of traditional search query.

NotebookLM uses natural language processing to make sense of your documents. When you type in a query, the system understands what you’re looking for. When I queried about the death of Mahler’s loved ones, I didn’t have to mention their names or even their relationship to him — NotebookLM understood what I was asking. These exploratory searches are more powerful than old-fashioned keyword searches, which only work if an exact word combination appears in your document. NotebookLM makes it easy to run abstract queries as well, searching for moments of anger or surprise.

Tip: target specific sources. You can use the checkboxes next to each source to limit your search to particular documents. This precision is handy when you want to search within a specific report or compare information across just two or three key documents.

Click here to get my FREE Claude course (20+ lessons)

Visualize information

Use the Studio tab to create shareable reports, slides, graphics, and multimedia out of your notebook material. Unlike other AI tools, NotebookLM’s creations are grounded in your source documents — they don’t pull from the Web or generic training data. Because they draw only from your source material, the creations will change as you add more to your notebook, or if you mark only a subset of sources to be used.

Create a mind map first to get an overview of the topics covered in a notebook. Then create the following elements to understand and share your material.

Infographics

Create polished visual summaries. Choose whether you want a landscape, portrait, or square image, and how simple or detailed it should be. Then type in an optional custom prompt to guide the design. You can include instructions about your preferred color palette, target audience, illustration style, and the kinds of numbers or facts to prioritize.

A caveat: NotebookLM consistently produces clean, readable text. It’s mostly accurate, but I’ve encountered occasional errors. Here’s an example: Mahler’s age of death is wrong at the bottom of this NotebookLM infographic.

Slide decks

NotebookLM’s newest capability — generating slide decks — continues to surprise me. When I ask it to make slides summing up notebook material — it comes up with outstanding results, like this slide deck about Mahler.

You can choose between detailed standalone slides, or simpler TED-style presenter slides meant to accompany a verbal presentation. As with the infographic tool, you can just press the slide deck button to let NotebookLM decide what to generate. But you’ll get something more relevant to you if you write a prompt to guide the visual style and subject matter focus. The slides include a small NotebookLM watermark in the bottom right corner.

Here’s an example of a slide deck about NotebookLM I created with NotebookLM

A caveat: In my testing, the slides have been clean and visually engaging. They’re not perfect, though. A deck about our new bilingual journalism program, for example, included misleading AI-generated images of our faculty members.

Subscribe now

Video overviews

Create a video summary of the material in your notebook. Think of it as an AI-narrated slide show. Fortunately, there’s no talking avatar. I like how these videos include facts, examples, quotes, and images pulled directly from your source documents.

Choose between a brief video (1-2 minutes) or a longer explainer (often six to 10 minutes). You can’t specify the exact length. Tailor the approach to your viewers with a prompt. You can even specify a specific audience, whether board members of a charity you’re presenting to, or grandchildren new to your subject matter.

Videos can take five to 10 minutes to generate. Free accounts can generate only a few videos, slide decks, or infographics per notebook before hitting a usage limit. When your video — or other creation — is ready, you can download and share it, or view it within your notebook.

📺 Here’s a video overview of NotebookLM I created with NotebookLM

Podcasts

NotebookLM’s audio overviews became Internet famous for their remarkably human-sounding conversations. When I played a clip for a group of students when this feature launched, they didn’t realize the speakers weren’t human.

Example: Here’s a new “Deep Dive” audio piece I generated about NotebookLM for this post.

• You can write a brief or detailed prompt to guide the style of the audio, and you can choose from multiple formats.

• After a few minutes, the audio file is ready for you to download and share.

• Tip: add an AI-generated label to this kind of audio or any other material you create with NotebookLM. That way people will know where it came from and won’t assume you created each detail from scratch.

You can generate audio pieces from a subset of your documents or your full collection of sources. Here are the four kinds of audio you can generate, with an example of each:

• Debate. Here’s an audio debate I prompted NotebookLM to create about which of its features are most useful.

• Critique. Here’s a critique of NotebookLM I generated from 19 sources I added.

• Brief summary. Here’s a 90-second audio overview.

• Deep dive. Here’s a deep dive NotebookLM explainer.

Text reports

In addition to multimedia, you can generate custom reports. The reports tend to be around 2,000 to 3,000 words, or six to 12 pages. Here are example reports generated by NotebookLM: an advanced guide to NotebookLM and a guide to integrating NotebookLM in a newsroom.

I’ve found the dozens of reports I’ve generated to be thorough enough to be useful for reference or learning. They also help point to sources worth exploring further. Try prompting NotebookLM to create the following kinds of reports:

• Timelines: Organize chronological information

• FAQs: Common questions and answers about your topic

• Explainers: Break down complex concepts

• Teaching guides: Useful if you’re an educator or lead workshops

• Student handbooks: Supplemental resources

• Critiques: Analysis of weaknesses or limitations in your sources

• Debate reports: Multiple perspectives on controversial topics

Flashcards and quizzes

When learning something new, create flashcards or quizzes with multiple-choice questions to test yourself.

• Describe your level of understanding (e.g. “I’m new to this,” or “I’m a professional in this field, but I’m new to this framework,”).

• Choose whether you want small or large number of questions or flashcards.

• Specify concepts you want the quiz or flashcards to focus on.

• You can also ask NotebookLM to focus on a particular source, like a certain link, PDF, or video you’ve uploaded. Example: Check out my NotebookLM flashcards.

5 Projects to Try

1. Organize a work project

Each time you add a file, NotebookLM summarizes it. Its full text is then searchable with citations, so you know you’re not getting AI hallucinations. To assemble a useful notebook, gather relevant documents, including:

• Plans, internal reports, or project memos

• Links to relevant sites

• Meeting recordings or transcripts

• Important emails copied and pasted or saved as PDFs or docs

• Background reports, company manuals, or competitive research

Use your project notebook to:

• Create summary reports or timelines to onboard new team members

• Draft slide decks for internal meetings

• Make infographics to visually summarize complex processes or workflows

• Quickly find relevant quotes, stats, anecdotes, or examples

• Refresh your memory when returning to the project later on

2. Plan a trip

I create travel notebooks to help me find relevant family activities and ideas for outings. I’ve done this before with Perplexity and other AI platforms, but I like the way NotebookLM lets me gather so many different kinds of inputs: links, videos, articles, and local guides—everything I might want to reference when planning weekend activities or hosting visitors. You can find these kinds of resources with a Google or Perplexity search, or do the whole process within NotebookLM.

For travel planning, compile these materials:

• Historical and cultural information

• Entertainment guides and reviews

• Restaurant recommendations

• Local blog posts, event listings, or links to top attractions

Then ask NotebookLM to generate:

• Itineraries

• FAQs about your destination

• Recommendations based on your budget or other constraints

• Slide decks or infographics to share with your travel companions

• Flashcards for learning key phrases if you’re traveling abroad

• Quiz games to play at the airport while waiting in line

3. Learn something

Here’s a meta use-case: I created a notebook about NotebookLM to help me learn about its nooks and crannies. (Try the quiz about NotebookLM it created for me.) I made another one about “deliberative dialogue” to learn more about tactics for encouraging civil discourse between people who violently disagree.

To build a learning notebook:

• Upload relevant YouTube videos, articles, and course materials.

• Use the “Add Sources” panel to add docs from your Google Drive or the Web.

• Generate mind maps, quizzes, and flashcards to test your understanding.

• Create audio guides to learn while exercising, cleaning, or commuting.

• Prompt for timelines, FAQs, explainers, infographics, and slide decks tailored to your knowledge level and learning goals.

Tip: break large documents into smaller pieces

NotebookLM uses retrieval augmented generation (RAG) for search. That keeps it grounded in your material and avoids hallucination. But it also means that when asked to quickly search gigantic documents, NotebookLM may have the capacity to scour only a subset of your source material.

To avoid searches that miss important material, consider breaking enormous documents into smaller pieces and narrowing your searches to specific sources or more precise subjects.

Subscribe now

4. Compile reference guides

Build notebooks to help you handle recurring tasks.

• Grant writing. Compile successful applications, guidelines, or evaluation criteria.

• Social posts. Gather style guides, brand guidelines, and examples of past posts that have worked well for you or competitors.

• Technical documentation. Assemble specs, organizational rules, or industry best practices.

• Customer research. Add past surveys, interview transcripts, analytics reports, or testimonials. Tip: as a first step, strip names and emails from surveys or interviews to protect respondents’ privacy.

5. Manage home projects

Create notebooks for life outside of work. NotebookLM is great for this because unlike other AI tools, it lets you input so many different kinds of sources with huge file sizes, whether you have videos, audio files, PDFs, your own handwritten notes, links to various sites, or Google Drive files.

• Recipe collections and guides to various cooking techniques

• Home improvement projects with how-to articles and product reviews

• Hobby research for woodworking, guitar, photography, or gardening

Why NotebookLM Stands Out 📍

• Unlike AI assistants designed around an open-ended chat box, NotebookLM is structured around a more familiar paradigm: a searchable notebook. The closest parallels are Claude Projects or ChatGPT Projects, which allow you to organize documents in a folder that can inform AI queries on those services. Perplexity Spaces is also useful for organizing related search threads. But none of those can generate NotebookLM’s full range of outputs, and each draws on its own training data as well as your sources.

• NotebookLM’s citation system means you can trust its search results, because you can see the cited section in your original document. And it’s unique in being able to generate everything from audio and video to reports, slides and infographics from your source materials.

  • Note: Citations aren’t provided within infographics, slide decks, video or audio overviews. If there are tidbits from those you want to trace back to a source, summarize the fact or detail in question in NotebookLM’s explore tab — the chat window — to ask for a citation.

• The free tier is powerful enough for most people. And it keeps improving, adding significant new capabilities every couple of months.

• The bottom line: if I were forced to recommend a single AI tool for many different kinds of readers, I’d pick NotebookLM.

More NotebookLM Guides

For more, read these guides about NotebookLM:

• Connect Claude to NotebookLM

• How to build your AI mentor with NotebookLM

• How we learn multiple subjects with NotebookLM

Leave a comment

Don’t forget to subscribe to Wonder Tools, where you can discover the Web’s most useful sites.

Join my newsletter with 70K+ people to master AI in 2026

Let’s connect! You can find me on LinkedIn, Instagram and X

Click here to get my FREE Claude course (20+ lessons)

Most people think they get bad results from AI because they write bad prompts.

I used to think the same.

So I spent months collecting prompt tricks, magic words, and those giant “perfect prompt” templates people love to share.

Some of them helped.

But my results were still hit or miss, and I couldn’t figure out why.

Then it clicked. The problem wasn’t the prompts. The problem was that AI kept answering before it actually understood what I wanted.

Think about it.

If you ask a coworker to write a report and they start typing two seconds later, without a single question, you know it’ll miss the point.

A fast answer with no context is usually an average answer.

AI works the same way.

When you give it an empty brief, it fills the gaps with guesses. When AI guesses, you get generic output.

The good news is you can fix this without memorizing another prompt formula.

You only need to change one thing: make AI understand the task before it starts working on it.

Here’s what we’ll cover:

• The Briefing Loop, the exact system I use to get better first drafts

• Real examples for decision-making, writing, and more

• How to save this context so you don’t have to repeat yourself every time

Subscribe now

The Problem: AI Responds Too Fast

One of the biggest problems with AI is that it answers even when it has no clue what you actually want.

If you ask it to translate a sentence, fix a typo, or reformat a table, a fast answer is totally fine, but if you ask it to write an article, prepare a class, analyze a dataset, design a proposal, or help you make a decision, a fast answer is usually a bad sign.

Here quality doesn’t come from the model.

It comes from context.

And when the context is missing, AI just fills the gaps with average answers. That’s why so many outputs sound the same. Not because everyone uses the same tool, but because everyone feeds it the same empty instruction.

So what we want is simple. AI should start working only after it understands the task, not before. This is what this guide is about.

Prompting Is Asking. Briefing Is Preparing

A prompt says:

Do this.

A brief says something like:

- This is the goal.
- This is who it's for.
- This is the tone.
- These are the limits.
- This is what I want to avoid.
- This is what a good result looks like to me.

Most people try to get better results by writing better prompts.

That helps, but only up to a point.

The real jump comes from somewhere else. It’s about giving direction and purpose to what you’re building, not just cleaner-sounding instructions.

So the clearer the brief, the less the model has to guess.

AI’s First Job Shouldn’t Always Be to Answer

This is the most important shift.

That shift changes the whole conversation.

AI stops acting like a machine that spits out text and starts working more like an editor, a consultant, a teacher, or a sparring partner.

Sometimes the best first response from AI isn’t an answer.

It’s a question.

Or even better, a short list of questions that show you what you haven’t made clear yet.

Picture this. You tell it:

I want to write an article about productivity

A bad interaction is AI writing the article right away.

A better one is AI asking you first:

- Who is the article for?
- What problem do you want to solve?
- What common idea do you want to challenge?
- What examples do you want to include?
- What should the reader be able to do after reading it?

That already makes the article better before you write a single line.

Subscribe now

The Solution: The Briefing Loop

Here’s the system.

It’s not complicated. But it completely changed the way I used to work with AI.

The goal is to stop AI from answering fast but badly.

Step 1: Start With a Rough Task

You don’t need to start with a perfect prompt.

That’s actually the whole point. You can start with something simple:

I want to write an article about why people get poor results from AI

This isn't a brief yet. It's just the rough task. It tells AI what you're trying to do, but it doesn't give it enough context to do it well.

That's exactly why the next step matters.

Step 2: Ask AI to Build the Brief Before Answering

Don’t ask AI to do the task yet. Ask it to turn your rough task into a proper brief first.

Use something like this:

Before you answer, turn my rough request into a proper brief.
Identify what context is missing or ambiguous.
Ask the fewest questions possible.
Only ask questions whose answers would change the final output.
If a question is nice to know but not necessary, skip it.
Ask no more than 3 questions for small tasks and no more than 5 questions for complex tasks.
Don't generate the final output yet.

This forces AI to pause.

And instead of rushing to finish the task with half the information, it helps you clarify the work first.

So before AI responds, you check:

This checklist helps you reduce the guessing and, as a result, get better outputs.

After putting together the prompts from steps 1 and 2 and running them, expect some back-and-forth before Claude fully understands what you want. Once you’re both aligned, move on to step 3.

Step 3: Have It Restate the Brief

This step looks small, but it saves you from a lot of mistakes.

Before generating the final output, ask:

Before writing, restate the brief in your own words.

Include:

- Goal
- Audience
- Angle
- Constraints
- Expected format
- What to avoid

Why does this work?

Because if AI got the brief wrong, it will also get the result wrong. And it’s way easier to fix a five-line restatement than to fix a whole draft.

Step 4: Only Then Generate the First Version

Now it’s ready.

AI knows what it has to do.

It knows who it’s for.
It knows what to avoid.
It knows what format to follow.
It knows what “good” looks like.

That doesn’t guarantee a perfect answer. But it makes the first version a lot stronger.

And a strong first version matters more than it seems. Because when the first version is weak, everything after it turns into patches. This way, you save all the time you'd otherwise spend fixing the mess a bad brief creates.

Optional: Run a Gap Review

Once you get the first result, don't ask:

Can you improve it?

That’s too limited.

Ask instead:

Review this result against the brief.

Tell me:

1. What’s strong
2. What’s still vague
3. What’s missing
4. What should be removed
5. What would make it more useful for the audience

This way AI doesn't just produce. It also becomes an evaluator.

Click here to get my FREE Claude course (20+ lessons)

Putting It Into Practice

Example 1: Writing an Article

Weak approach:

Write me an article about using AI better.

Better approach:

I want to write an article for professionals who use AI every day but still get generic answers. The core idea is that the problem isn't bad prompts, it's bad briefs. 
Before writing, review this brief. Ask me only the 5 most important questions needed to make the article specific, useful, and not generic.

AI might ask:

• Who exactly is the reader?

• What kinds of tasks do they use AI for?

• What belief do you want to challenge?

• What examples do you want to include?

• What should the reader do differently after reading?

That already makes the article better before you write a single line.

Example 2: Teaching a Complex Topic

Weak approach:

Explain confidence intervals to me

Better approach:

I need to explain confidence intervals to students who are learning statistical inference for the first time. Before explaining, ask me what you need to know about their current level, the notation they know, the examples they've already seen, the mistakes they usually make, and whether the explanation should be intuitive, formal, or both. Don't explain anything yet.

This completely changes the answer. Because "explain confidence intervals to me" can mean a lot of different things: a simple analogy, a formal derivation, an exam solution, or a script for a 10-minute class.

Example 3: Data Analysis

Weak approach:

Analyze this dataset

Better approach:

I'm going to analyze a dataset for a business report. 
Before suggesting any analysis, ask me the essential questions about the business objective, the report's audience, the available variables, the decision the analysis should support, the expected technical level, and the final deliverable format. Don't start the analysis until the brief is clear

Because with AI you have to accept one thing: "analyze the dataset" isn't a goal. Analyze it for what? To predict, compare groups, find patterns, detect anomalies, make a decision, build a dashboard, or train a model? Each one points to a different analysis.

Example 4: Making a Decision

Weak approach:

Should I do this?

Better approach:

I'm evaluating this decision: [describe the decision]. 
Before recommending anything, ask me the minimum number of questions needed to understand my goal, my constraints, the alternatives, the risk of being wrong, the cost of waiting, and what information would change the decision. Then give me a recommendation with trade-offs

Here, the method stops being just about productivity. It turns into a way of thinking and getting clarity, because the right questions:

• Force you to clarify what you actually want.

• Reveal constraints you never mentioned.

• Surface risks you’ve been ignoring.

• Help you separate a preference from a real decision.

One More Thing

You don’t need to use the Briefing Loop for everything.

Use it when the task involves ambiguity, judgment, or real consequences.

The key question is this:

Does AI need context to do this well?

If the answer is no, just ask directly.

And one last thing.

If you keep explaining who you are, how you write, who your audience is, what tone you use, or what kind of results you expect, you’re basically writing the same brief over and over by hand.

That’s where context files come in.

A context file is a permanent brief. You write it once, and AI reuses it every time.

I just released My AI System. It contains a skill that generates context files tailored to your needs. Check it out: artificialcorner.com/p/ai-system

If this helped, subscribe to Artificial Corner. I share practical AI guides, Claude workflows, and systems you can actually use the same day you read them.

Join my newsletter with 70K+ people to master AI in 2026

Let’s connect! You can find me on LinkedIn, Instagram and X

Subscribe now

If you’re new here, welcome! I’m Frank. My online writing journey started in 2021, when I began publishing on Medium(dot)com. I grew to 150,000 followers there and then moved to Substack in 2023 where I share weekly AI guides.

You’re reading this because you like the way I write.

The Claude guides, the emails, the posts.

They’re easy to follow. Sounds like a friend teaching you what he’s learned, not a robot.

But here’s what you might not know.

Claude writes the first draft. Not from a generic prompt. Not from “write in my style.” I built a system around my voice.

For every project, Claude has access to files that shape the writing. These files teach Claude how I write: the words I like/dislike, the analogies I like, the punctuation I hate, the structure my guides follow, and more.

After Claude generates the first draft, I do the final pass: I add personal knowledge, make the visuals, delete what I disagree with, sharpen the examples, etc.

I make it feel human.

So the system doesn’t replace me.

It gives me a strong first draft in my voice (fast) and lets me spend my energy on the parts that actually matter. That’s the Claude system behind the writing 70,000+ of you read here. And the system I’ll share with you today.

Not as a “copy my newsletter” template (most of you don’t run a newsletter).

The backbone.

The structure you can adapt to your own projects and workflow.

All in your voice.

But before I hand it over, let me show you where it lives and what the folder contains.

Why this system lives in Claude Cowork

Over the past three years, I’ve built and run my business with AI at the center. Along the way, I’ve relied on three main tools: ChatGPT, Claude Cowork, and Claude Code.

Each one taught me something.

ChatGPT is the easiest to start with, but you’re stuck in a chat bubble. You ask, you copy, you paste, you fix. It never touches your real files or remembers how you work.

Claude Code is the most powerful AI I’ve ever used. There’s just one catch: the more power you want to unlock, the more technical you have to get. For most of you, that’s a wall.

Claude Cowork sits right in the sweet spot.

• Cowork is as easy to use as ChatGPT

• Cowork has a big chunk of the power of Claude Code

• No coding. No terminal. No complex setup

A real system used to mean going technical.

Cowork killed that trade-off: you can ask Cowork to generate drafts, customize the system, and develop it using plain English instructions (the same way you’d ask a coworker for a hand).

So that's where I built this. You get the system, not just prompts.

What’s in the AI system folder

Here’s what the AI system folder contains:

• Context files: they teach Claude your writing voice and who you write to

• Claude skills: the core skill that generates your context files, plus two additional skills that turn videos and articles into usable source material for Claude

This is the foundation of the system.

For this first version, I kept only the essentials: the pieces anyone can adapt to their own projects (the context files). I’ll keep adding new Claude skills, new files, and new upgrades that plug into the same system in the coming months.

That said, this first folder has everything you need to make Claude write in your voice for whatever project you have in mind.

Download my Cowork writing system

This is the core system behind everything I publish.

The only reason I can keep building it (and keep the free guides free) is the support of paid subscribers. Become one today and it's yours!

All paid subscribers get:

• The full kit: the context profiles and Claude skills

• Monthly updates: New Claude skills, project files, and more to keep the system growing

In addition to that, annual subscribers get (only this week):

• voice-dna-frank.md: In case you want to skip the setup and write your drafts with my voice

By the way, this month Artificial Corner turns 3 years old.

The first 20 people who go annual get a 20% discount

Click here for the discount: https://artificialcorner.com/endsunday

Subscribe now

🔒 Paid Substack members: link to get the files below

Read more

Click here to get my FREE Claude course (20+ lessons)

For a long time, I used AI the same way many people do.

I’d ask a question, copy the answer, and move on.

It felt productive. I was saving time, getting things done faster, and checking more boxes. But at some point, I realized something uncomfortable: I was getting answers faster, but I wasn’t thinking any better.

AI was doing the work, and I was just collecting the output.

Now, instead of treating AI like a vending machine for answers, I use it to improve my inputs, learn faster, challenge my decisions, and free up time for the work that actually needs me.

The difference was bigger than I expected.

I wasn’t just faster. I was sharper.

In this guide, I’ll show you how to make that shift.

If you’re getting value from my AI guides, support us by becoming a paid subscriber👇

Subscribe now

1) The real shift: using AI as a thinking partner

Most people use AI like a calculator. You type a question, an answer pops out, problem solved.

But AI doesn’t work like a calculator.

A calculator gives you precise, consistent results every time. AI can be brilliant, fast, and creative… but it can also sound super convincing while being completely wrong (and that’s the tricky part).

So the worst way to use AI is to treat it like a vending machine for answers. You put a coin in, you grab whatever falls out.

The better way? Treat it as a thinking partner.

Don’t just measure how much time AI saved you. Measure how much better you were able to think because of it.

Here’s what that shift looks like:

2) Better inputs, better ideas

AI can speed up your results. No doubt.

But if your inputs are poor, your outputs will be poor too. They’ll just arrive faster.

Before you obsess over prompts, take a look at what you feed your brain every day. If your attention is scattered across shallow info, random content, and updates that add nothing to your work or learning, AI won’t fix that for you. At best, it’ll help you summarize the mess a little faster.

The smart move is the opposite. Use AI to improve the quality of what enters your mind, instead of drowning yourself in trivial stuff.

Here are three strategies you can start using today:

a) Use AI to curate information, not to flood you with more

One of the most useful things you can do is build a daily briefing around the topics that actually matter in your field.

Instead of opening ten tabs, half-reading headlines, and slipping into autopilot, you can ask AI to do the first scan, pull out the essentials, and hand you only what matters.

Here’s a prompt I use a lot:

Act as my research analyst. Find the three most important developments about [topic] in the last 24 hours. Summarize each one in two sentences, explain why it matters, and tell me what strategic question I should be asking myself because of it.

And you get something like this:

I know at first glance this looks like just another prompt. But what I’m really doing is cutting the noise without cutting the thinking. I want AI to give me context, not to think for me.

b) Don’t learn “just in case.” Learn “just in time”

Another common mistake is hoarding content you never actually use.

It looks like this:

Learning this way feels responsible, but most of the time it’s just an illusion of progress. You read a lot, apply little, and remember even less (we’ve all been there).

So flip it. Learn the thing right before you need to use it.

A couple of examples:

• You have a pricing meeting tomorrow → learn what you need about pricing a few days before

• You’re designing an executive presentation next week → study how to structure executive arguments now.

When learning is tied to an action that’s right in front of you, it sticks way better. AI just shortens the distance between “I read it” and “I used it.

Subscribe now

3) Use AI as a tutor

When AI hands you the answer too fast, it feels efficient. But feeling efficient and actually learning are two different things.

Learning isn’t just getting exposed to information.

Learning means putting in effort, recalling stuff, messing up, adjusting, and trying again.

So if you want to use AI to get smarter, you have to change the job you give it.

Start using AI as a tutor. Here’s how to think about it:

a) The problem with getting the solution too fast

Really understanding something means you can explain it in your own words, use the idea in a different situation, answer new questions about it, and catch the holes in your own reasoning.

For that to happen, AI can’t do all the work. It has to make you show up too.

b) The best setup: you try first, then AI corrects you

Here's a really powerful way to use it:

For example, one prompt I use:

I just studied this topic: [topic]. Ask me questions one by one. Don't give me the answer until I try to respond. After each answer, tell me what I got right, what I missed, and then ask me a harder question.

This turns AI into a tutor. It walks you through the whole process of locking in what you just learned (and it's surprisingly good at it).

c) Turn friction into part of the learning process

There’s a simple idea here. For information tasks, reduce friction. For learning tasks, keep some of it.

In plain terms:

• If you want to catch up fast → let AI be fast.

• If you want to truly learn → let it make you work for it.

So under that logic, you can ask AI:

The real advantage of AI here isn’t that it “knows a lot.”

The advantage is that it’s always there, it never gets tired of explaining the same thing, and it can adjust to your level on the spot (try getting that from a textbook).

Click here to get my FREE Claude course (20+ lessons)

4) Make AI attack your ideas before reality does

Almost every bad decision has the same thing behind it: someone too close to their own idea.

When you’re the one who came up with it, you see the potential, the excitement, the promise. What you don’t see as easily are the cracks.

That’s where AI gets really valuable. Not just to help you think an idea through, but to rip it apart on purpose.

a) The most common mistake: using AI to get a pat on the back

Most people drop a plan into AI and ask, “What do you think?”

The problem is that if you don’t steer it, AI tends to play nice. It’ll summarize, validate, and tidy things up, but it won’t push back with the rigor you actually need.

And if all it does is hand you a prettier version of what you already wanted to believe, it’s not helping you think. It’s just agreeing with better grammar.

So ask for opposition, not approval.

Here's the skill I use:

## Sycophancy Skill

You are my critical thinking partner. Your default mode is constructive disagreement.

1. Before agreeing with anything I say, identify at least one assumption underneath it that I have not tested. State the assumption plainly.
2. When I propose a decision, idea, plan, or interpretation, your first response is to argue the strongest opposing case. Do not soften it. Do not append “but you might be right.” Make me defend my position.
3. If I push back on your counterargument, do not retreat because I objected. Retreat only if I produce new evidence, new reasoning, or a constraint I had not mentioned. Saying “fair point” without new information is not enough.
4. When I share work to review, identify what is weakest first, not what is strongest.
5. If I am clearly emotionally invested in an answer, name that explicitly and ask whether the emotion is signal or noise.
6. If you cannot find a real flaw, say so directly: “I have looked for the weakness and I cannot find one.” Do not invent a flaw to perform thoroughness.
7. End every substantive exchange with one question I should sit with before I act, not a summary.

Tone: direct, not aggressive. Specific, not abstract. One disagreement at a time. Cite my own words when challenging me. Do not open with praise, do not use “great question” or any flattery, do not hedge with “I could be wrong but,” and do not add a closing reassurance.

You can also use prompts to turn AI into a sparring partner.

Here are 3 good prompts:

Prompt #1:
Imagine this plan fails six months from now. Walk backward and tell me the most likely reasons why it failed.

Prompt #2:
Act as a smart, cynical competitor. Analyze this plan and tell me how you'd exploit its weaknesses.

Prompt #3:
Rank the three main risks of this plan by probability and impact. Then give me one preventive action for each.

b) The real value: protecting you from your own blind spots

What makes this so useful is that it doesn’t even need AI to be perfectly right. It forces you to look where you normally wouldn’t, and that alone sharpens your judgment.

So AI becomes a kind of honest mirror, as long as you tell it to stop nodding and start arguing.

Before any big decision, you can run this map:

5. A simple system to use AI better every week

Getting smarter with AI doesn’t come from finding the perfect prompt.

It comes from running a few thinking loops on repeat.

You can start with something simple:

a) Every day: improve your inputs

Use AI to get a short briefing on your field. The goal isn't more information. It's better information.

Prompt:

Give me the three most important things I should understand today about [topic]. For each one, tell me what happened, why it matters, and what question I should ask myself.

b) Every time you learn something: turn it into practice

Don't stop at the explanation. Push it to quiz you.

Prompt:

Ask me questions about this [topic] one at a time. Don't give me the answer until I've tried. Then correct my reasoning and crank up the difficulty.

c) Before an important decision: ask for opposition

Skip the validation. What you want here is resistance.

Prompt:

Before you agree with me, point out the main weakness of this idea and the most dangerous assumption I'm making.

d) Before creating something: improve the brief

Resist the urge to ask for the final output right away. Clarify the problem first.

Prompt:

Before writing, tell me what you think I'm trying to achieve, who the audience is, what tone makes sense, and what info is missing to do it right.

Then, at the end of each week, look back at what you delegated and what you protected. A few questions worth asking yourself:

• What repetitive tasks could I have handed off?

• Where did I use AI to think less?

• Where did I use AI to think better?

• What decision got better because I questioned it first?

• What part of my work really needed my own judgment?

One more thing

Now that you’ve seen the whole framework, don’t just nod and close the tab.

Build your own weekly system for thinking better with AI:

I wrote this guide to help you use AI to get smarter.

Becoming a paid subscriber gets you my full library of guides, the prompts and systems I actually use, and new ones every week.

Subscribe now

Join my newsletter with 70K+ people to master AI in 2026

Let’s connect! You can find me on LinkedIn, Instagram and X

Click here to get my FREE Claude course (20+ lessons)

Claude isn’t the bottleneck. Your computer is.

You can have the best prompts in the world, but if you’re fighting your computer every time you switch windows, snap a layout, paste a screenshot, attach a file, or keep your screen from dimming on a long Cowork run, Claude is going to feel harder to use.

The apps below remove the friction around Claude.

In this guide, you’ll learn:

• The 8 Mac apps that pair best with Claude (+ Windows alternatives)

• The exact use case that earns each one a spot on your dock

• How I make the most out of these apps

Subscribe now

Read more

Click here to get my FREE Claude course (20+ lessons)

I used to blame my prompts when Claude gave me mediocre results.

Turns out, the problem was sitting in my folder the whole time.

Files named final.docx, final_v2.docx, new_final_REAL.docx. Old notes mixed with current drafts. No clear structure, no clear order.

Claude was doing its best with a messy workspace. And messy workspace = messy output.

A clean file system changes that. Claude knows what to read, what to use, and where to save the result. The output gets better without changing the prompt at all.

In this guide, I’ll show you the system I built: five folders, simple naming rules, and a structure Claude (and you) can actually navigate.

Here’s what we’ll cover:

• Why your folder is now part of the prompt (and what happens when it’s a mess)

• The five-folder structure that gives Claude a clear map of your workspace

• How to name files so Claude knows what’s inside before opening them

• The difference between active work, reusable context, and old files

• The full system, all in one view

If you’re getting value from my Claude guides, support us by becoming a paid subscriber👇

Subscribe now

1. Your folder is now part of the prompt

A prompt isn’t just the text you type.

It’s also everything in your folder:

That’s why a messy folder gives Claude messy context.

It might open the wrong draft. Treat an old note as if it were current. Or ignore the file that actually matters. Instead of doing the work, it ends up spending time trying to figure out your workspace.

A well-organized folder removes that guesswork. It tells Claude where things are.

The goal is to make the folders easy to understand. For you and for Claude.

2. Build a workspace Claude can navigate

You don’t need a new system.

Your current Claude system just needs to handle more files without turning into noise.

The foundation stays simple:

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

But when Claude starts working with many files, two small adjustments help:

1. Number the first level of the workspace

2. Add archive/

YOUR-WORKSPACE/
├── [01] system/
├── [02] context/
├── [03] projects/
├── [04] outputs/
└── [99] archive/

I use [99] archive/ so it always stays at the end. It’s not active work, but it stays searchable.

Same workspace. Easier to navigate.

[01] system/ → rules that Claude must follow
[02] context/ → stable context that Claude must know
[03] projects/ → active work
[04] outputs/ → completed work
[99] archive/ → old, but searchable

That’s enough. You don’t need twenty folders or a perfect taxonomy.

You need a structure Claude can scan quickly and that is:

Subscribe now

3. Name files so Claude (and you) can understand before opening them

File names are context.

Before Claude opens a document, the name already tells it what that file might be.

Avoid this:

• final.docx

• final_v2.docx

• new_final.docx

• final_REAL.docx

• notes.md

• screenshot.png

Use names that answer three questions: When? What? What status is it in?

Let’s look at an example:

Not every file needs the same level of time precision. Use the level that matches the file type:

A few quick rules:

1. YEAR → best for long-term files

2. YEAR + QUARTER → best for planning cycles

3. YEAR + MONTH → best for recurring work

4. FULL DATE → best for task-specific files

Click here to get my FREE Claude course (20+ lessons)

4. Separate active work, reusable knowledge, and old context

Not all files need the same treatment.

Some are active.
Some are reusable.
Some are already finished.
Some are old, but might be useful later.

Don’t mix them.

• projects/ → active work

• context/ → stable context and reusable references

• outputs/ → finished work

• archive/ → old, but searchable

Numbering helps when you need a clear path. Claude knows where to start, which folder has the rules, where the active work is, and where the results go.

That doesn’t mean every file needs a number, though.

Inside [03] projects/, numbering makes sense because there’s a workflow:

[02] Projects/ 
└── claude-file-system-article/ 
├── 00_BRIEF.md 
├── 01_SOURCES/ 
├── 02_NOTES/ 
├── 03_DRAFTS/ 
├── 04_VISUALS/ 
└── 05_FINAL/

That order tells Claude:

[02] context/ works differently. Think of it as a library, not a workflow. There's no universal reading order here.

[02] context/ 
├── HOW-I-WORK.md 
├── voice-guide.md 
├── audience.md 
└── references/   
  ├── brand-guide.md 
  ├── article-examples/ 
  ├── prompt-library/ 
  └── source-library/

Reusable files don't need to be copied into every project.

Just tell Claude to use them when needed:

• Reusable across projects → context/

• Specific to one project → projects/

• Finished work → outputs/

• Old but useful later → archive/

Here’s the rule I follow for each folder:

Workspace → numbered map
Projects → numbered workflow
Context → descriptive library
Outputs → date + topic + status
Archive → old but searchable

Subscribe now

5. This is how everything looks together

The complete system doesn’t need to look complicated.

It just needs to show Claude where everything is.

YOUR-WORKSPACE/
├── [01] system/
│   └── SYSTEM-RULES.md
│
├── [02] context/
│   ├── HOW-I-WORK.md
│   ├── voice-guide.md
│   ├── audience.md
│   └── references/
│       ├── brand-guide.md
│       ├── article-examples/
│       └── source-library/
│
├── [03] projects/
│   └── claude-file-system-article/
│       ├── 00_BRIEF.md
│       ├── 01_SOURCES/
│       │   ├── 2026-06-24_video-notes.md
│       │   └── 2026-06-24_cowork-workspace-notes.md
│       ├── 02_NOTES/
│       ├── 03_DRAFTS/
│       │   └── 2026-06-24_claude-file-system_draft-v1.md
│       └── 04_VISUALS/
│           ├── workspace-map.png
│           └── file-name-structure.png
│
├── [04] outputs/
│   └── 2026-06-24_claude-file-system_final.md
│
└── [99] archive/
└── old-projects/

Don’t memorize the structure.

Memorize the rule:

• Markdown for instructions.

• Numbers for order.

• Names for meaning.

• Dates for search.

• Archive for noise.

That’s all.

Claude doesn’t need a perfect folder. Just one that tells it what to read, what to use, what to ignore, and where to save the result.

One more thing

Don’t reorganize your entire computer tonight.

That’s how people turn file management into another project they never finish.

Start with one folder Claude already touches.

Add structure.
Rename the files that matter.
Move the old stuff to [99] archive/.

That’s enough.

Give Claude fewer reasons to guess. That’s all this is.

If you want more practical guides like this, for Claude, Cowork, and AI workflows, the paid subscription is where you'll find them.

Subscribe now

Join my newsletter with 70K+ people to master AI in 2026

Let’s connect! You can find me on LinkedIn, Instagram and X