This is cool, thank you!

Some things I found from my own interactions across multiple models (in addition to above):

- It's basically all about the importance of (3). You need a feedback loop (we all do). and the best way is for it to change things and see the effects (ideally also against a good baseline like a test suite where it can roughly guage how close or far it is from the goal.) For assembly, a debugger/tracer works great (using batch-mode or scripts as models/tooling often choke on such interactivie TUI io).

- If it keeps missing the mark tell it to decorate the code with a file log recording all the info it needs to understand what's happening. Its analysis of such logs normally zeroes the solution pretty quickly, especially for complex tasks.

- If it's really struggling, tell it to sketch out a full plan in pseudocode, and explain why that will work, and analyze for any gotchas. Then to analayze the differences between the current implementation and the ideal it just worked out. This often helps get it unblocked.

> If there is anything Claude tends to repeatedly get wrong, not understand, or spend lots of tokens on, put it in your CLAUDE.md. Claude automatically reads this file and it’s a great way to avoid repeating yourself.

Sure, for 4/5 interactions then will ignore those completely :)

Try for yourself: add to CLAUDE.md an instruction to always refer to you as Mr. bcherny and it will stop very soon. Coincidentally at that point also loses tracks of all the other instructions.

Thank you for Claude Code (Web). Google has a similar offering with Google Jules. I got really, really bad results from Jules and was amazed by Claude Code when I finally discovered it.

I compared both with the same set of prompts and Claude Code seemed to be a senior expert developer and Jules, well don't know who be that bad ;-)

Anyway, I also wanted to have persistent information, so I don't have to feed Claude Code the same stuff over and over again. I was looking for similar functionality as Claude projects. But that's not available for Claude Code Web.

So, I asked Claude what would be a way of achieving pretty the same as projects, and it told me to put all information I wanted to share in a file with the filename:.clinerules. Claude told me I should put that file in the root of my repository.

So please help me, is your recommendation the correct way of doing this, or did Claude give the correct answer?

Maybe you can clear that up by explaining the difference between the two files?

Hey Boris,

I couldn't agree more. And using Plan mode was a major breakthrough for me. Speaking of Plan Mode...

I was previously using it repeatedly in sessions (and was getting great results). The most recent major release introduced this bug where it keeps referring back to the first plan you made in a session even when you're planning something else (https://github.com/anthropics/claude-code/issues/12505).

I find this bug incredibly confusing. Am I using Plan Mode in a really strange way? Because for me this is a showstopper bug–my core workflow is broken. I assume I'm using Claude Code abnormally otherwise this bug would be a bigger issue.

+1 on that Opus 4.5 is a game changer I have used to refactor and modernize one of my old react project using bootstrap, You have to be really precise when prompting and having solid CLAUDE.md works really well

Do you recommend having Claude dump your final plan into a document and having it execute from that piece by piece?

I feel like when I do plan mode (for CC and competing products), it seems good, but when I tell it to execute the output is not what we planned. I feel like I get slightly better results executing from a document in chunks (which of course necessitates building the iterative chunks into the plan).

> 1. If there is anything Claude tends to repeatedly get wrong, not understand, or spend lots of tokens on, put it in your CLAUDE.md.

What a joke. Claude regularly ignores the file. It is a toss up: we were playing a game at work to guess which items will it forget first: to run tests, formatter, linter etc. This is despite items saying ABSOLUTELY MUST, you HAVE To and so long.

I have cancelled my Claude Max subscription. At least Codex doesn’t tell me that broken tests are unrelated to its changes or complain that fixing 50 tests is too much work.

  > I add to my team’s CLAUDE.md multiple times a week.

also after you have a to-and-fro to course correct it on a task, run this self-reflection prompt

https://gist.github.com/a-c-m/f4cead5ca125d2eaad073dfd71efbc...

That will moves stuff that required manually clarifying back into the claude.md (or a useful subset you pick). It does a much better job of authoring claude.md than I do.

Thanks for your work great work on Claude Code!

One other feature with CLAUDE.md I’ve found useful is imports: prepending @ to a file name will force it to be imported into context. Otherwise, whether a file is read and loaded to context is dependent on tool use and planning by the agent (even with explicit instructions like “read file.txt”). Of course this means you have to be judicial with imports.

Hah, that's funny. Claude can't help but mess all the comments in the code up even if I explicitly tell it to not change any comments five times. That's literally the experience I had before opening this thread, never mind how often it completely ignores CLAUDE.md.

I would LOVE to use Opus 4.5, but it means I (a merely Pro peon) can work for maybe 30 minutes a day, instead of 60-90.

How do you make Claude code to choose opus and not sonnet? For me it seems to do it automatically

I’ve yet to see any real work get done with agents. Can you share examples or videos of real production level work getting done? Maybe in a tutorial format?

My current understanding is that it’s for demos and toy projects

Does the same happens if I create an AGENTS.md instead?

> I add to my team’s CLAUDE.md multiple times a week.

This concerns me because fighting tooling is not a positive thing. It’s very negative and indicates how immature everything is.

Hey Boris, can you teach CC how to use cd?

> Use Opus 4.5.

This drives up price faster than quality though. Also increases latency.

3. Puppeteer? Or Playwright? I haven't been able to make Puppeteer work for the past 8 weeks or so ("failed to reconnect"). Do you have a doc on this?

Does all my code get uploaded to the service?

Hey Boris from the Claude Code team - could you guys please be so kind so as to stop pushing that narrative about CLAUDE.md, either yourselves or through influencers and GenAI-grifters? The reason being, it is simply not true. A lot of the time the instructions will be ignored. Actually, the term "ignored" is putting the bar too high, because your tool does not intentionally "ignore", not having sentience and knowledge. We experience the effects of the instructions being ignored, because your software is not deterministic, its merely guessing the next token, and sometimes those instructions tacked onto the rest of the context statistically do not match what we as humans expect to see (while its perfectly logical for your machine learning text generator, based on the datasets it was trained on).