13 Comments
User's avatar
Marc Schärer's avatar

Thank you for the article.

Very informative.

It could be even better if it appropriately compared Claude Desktop with the Codex Desktop app which is ChatGPTs all in one desktop app offering (not coding like the cli).

That would help people more given the difference from codex desktop app and Claude desktop including workspace, dispatch, connector, skills etc is close to 0 these days.

Frank Andrade's avatar

True. I have to admit I haven't used Codex as much as Cowork. A few weeks ago, I started using Codex, though, because I needed an alternative to Cowork's computer use. I love this feature, but my issue with Cowork is that it takes full control of the computer, so you can't do anything while it's working.

Codex doesn't have this issue! Plus, I noticed it's more fluent than Cowork's computer use.

So far I've only used it for that, but, yes, I need to use it more to make a fairer comparison

Julia Musson's avatar

From Claude results to verify the accuracy of this post:

“One thing to flag: the title says “Claude” can do these things, but almost all 7 items are actually Claude Cowork features specifically — a desktop app feature requiring a paid plan (Pro/Max/Team/Enterprise) and macOS/Windows. If someone reads this and goes to the regular claude.ai chat or mobile app expecting Cowork, Skills, Scheduled Tasks, or Live Artifacts, they won’t find them there. That’s a meaningful blurring of “Claude” vs. “Claude Cowork” that the post doesn’t clearly disclose until you’re a few paragraphs in.”

The NotebookLM CLI workaround (point 6) involves a third-party GitHub repo I can’t independently verify works as described — treat that one with more skepticism than the others.

Sharon Seldner's avatar

Thanks for all the insight. I have spent the last 6 weeks on a project with another AI program. When I finish (almost there!) is there any merit to running it through Claud or Codex for an alternate outcome? I need to add charts for expansion “what ifs” and I would like some suggestions for other charts. There are probably hundreds, it’s a huge project with many divisions. Or is this a bad idea?

I don’t have time to start over but I also don’t want to miss an opportunity.

Julia Musson's avatar

IMHO:

AI/chat/LLM SYSTEMS of SYSTEMS may return different results ( whether GROK ChatGPT Gemini Claude )

Even the same prompt to the same system may yield different outcomes at different times.

Better minds than mine recommend:

1. querying for links, sources, and references for further research;

then

2. verifying search results against primary, original, as seminal sources as much as possible.

Another way to receive somewhat more accurate information is to tell your overly friendly chat systems:

✅ “If uncertain, say uncertain.”

✅ “Separate verified facts from inference.”

✅ “Provide primary sources and confidence level.”

✅ “State assumptions explicitly.”

✅ “Distinguish observation from recommendation.”

✅ “Cite original authoritative sources.”

✅ “Show evidence chain.”

Constraint-based prompting tends to work better than generic honesty requests:

https://www.techradar.com/ai-platforms-assistants/this-gemini-and-chatgpt-trick-will-stop-them-from-over-explaining-everything

Prompt sameness ≠ output sameness

According to the following in plain English website :

https://ai.plainenglish.io/stop-guessing-heres-what-happens-when-you-give-the-same-prompt-to-chatgpt-gemini-and-claude-da0a970ddf15

Models MAY differ because of:

* training distributions

* system instructions

* alignment policies

* context handling

* inference strategies

* safety tuning

* retrieval/tool behavior

Identical prompts CAN create:

* different recommendations

* different risk conclusions

* different prioritization

* different executive summaries.

Noe's avatar

Unrelated question, are monthly subscriptions cancellable and how? (Easy via Apple but wasn’t sure when done direct). Want to try for a month or twos

Frank Andrade's avatar

When done through Substack, you have to go to my Substack artificialcorner.com, click on your photo profile -> manage subscriptions. There, you can cancel

That said, I think you'll want to stay for more than one/two months 😅

The AI Runtime's avatar

Connectors and Skills are game changers

Frank Andrade's avatar

Yep! Skills is my favorite feature

Julia Musson's avatar

References provided by Copilot search:

Below is a verified, research‑ready framework for checking claims about Claude features, using the most authoritative sources available.

🧩 Core authoritative sources on Claude models

These are the primary references you will want for verification:

• Anthropic Model Cards

Official specs for Opus, Sonnet, Haiku, and Claude Code.

https://www.anthropic.com/news

• Anthropic API Documentation

Technical details on features, context windows, tools, memory, and Claude Code.

https://docs.anthropic.com

• Claude Code & Agent SDK Documentation

Covers code execution, file creation, agent workflows, and checkpoints.

https://docs.anthropic.com/claude/docs/claude-code (docs.anthropic.com in Bing)

https://docs.anthropic.com/claude/docs/agent-sdk (docs.anthropic.com in Bing)

• Anthropic Research Papers

Safety, alignment, and model architecture papers.

https://www.anthropic.com/research

• Benchmarks• SWE‑Bench Verified (coding)

https://www.swebench.com

• AIME (math)

https://aimecontest.org

• GPQA (reasoning)

https://github.com/Idavidrein/gpqa

• OSWorld (agentic computer use)

https://osworld.dev

Shivansh Gupta's avatar

ChatGPT Plus do have connectors, subagents to work code etc as well as automation schedulers too .

Linuxrebel's avatar

Great article. Only problem I have is that there's so much useful information. I don't know where to start. 😁

Doron Shachar's avatar

Excellent article.