5 ms·
> 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 th
by epolanski 11mo ago
> 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.
- dayjah 11mo agoThe Attention algo does that, it has a recency bias. Your observation is not necessarily indicative of Claude not loading CLAUDE.md. I think you may be observing context rot? How many back and forths are you into when you notice this?
- Fargren 11mo agoThat explains why it happens, but doesn't really help with the problem. The expectation I have as a pretty naive user, is that what is in the .md file should be permanently in the context. It's good to understand why this is not the case, but it's unintuitive and can lead to frustration. It's bad UX, if you ask me. I'm sure there are workarounds such as resetting the context, but the point is that god UX would mean such tricks are not needed.
- girvo 11mo agoYeah the current best approach to aggressively compact and recreate context by starting fresh. It’s awkward and I wish I didn’t have to.
- toxic72 11mo agoI'm surprised this hasn't been been automated yet but I'm pretty naive to the space - the problem of "when?"/"how often?" seems like a fun one to chew on
- evrenesat 11mo agoI think Gemini 3 pro (high) in Antigravity does something like that because I can keep asking for different changes in the same chat without needing to create a new session.
- tacitusarc 11mo agoIt’s not that it’s not in the context, it’s that it was injected so far back that it is deemed not so important when determining the next token.
- epolanski 11mo agoI know the reason, I just took the opportunity of answering to a claude dev to point out why it's no panacea and how this requires consistent context management. Real semi-productive workflow is really a "write plans in markdowns -> new chat -> implement few things -> update plans -> new chat, etc".
- SV_BubbleTime 11mo agoThe number of times I’ve written “read your own fucking Claude.md file” is a bit too numerous. “You’re absolutely right! I see here you don’t want me to break every coding convention you have specified for me!”
- theshrike79 11mo agoHow long are your conversations with Claude? I've used it pretty extensively over the year and never had issues with this. If you hit autocompact during a chat, it's already too long. You should've exported the relevant bits to a markdown file and reset context already.
- bryanrasmussen 11mo agohow often in an hour do you export the relevant bits to a markdown file and reset context?
- bcherny 11mo agoYeah, adherence is a hard problem. It should be feeling much better in newer models, especially Opus 4.5. I generally find that Opus listens to me the first time.
- frankdenbow 11mo agoHave been using Opus 4.5 and can confirm this is how it feels, it just works.
- PufPufPuf 11mo agoIt also works your wallet
- fastball 11mo agoIf you pay for a Claude Max subscription it is the same price as previous models.
- shepherdjerred 11mo agoJust wait a few months -- AI has been getting more affordable _very_ quickly
- wyre 11mo agoHighly recommend Claude Max, but I also want to point out Opus 4.5 is the cheapest Opus has ever been. (I just learned ChatGPT 5.2 Pro is $168/1mtok. Insanity.)
- radio879 11mo agoRight now Google Antigravity has free Claude Opus 4.5, with pretty decent allowances. I also use Github Copilot which is just $10/mo. I have to use the official copilot though, if I try to 'hack it' to work in Claude Code it burns thru all the credits too fast. I am having a LOT of great luck using Minimax M2 in Claude Code, its very cheap, and it works so good.. its close to Sonnet in Claude Code. I use this tool called cc-switch to swap out different models for Claude Code.
- 11mo ago
- roughly 11mo agoOne of the things you get an intuition for after using these systems is when to start a new conversation, and the basic rule of thumb is “always.” Use a conversation for one and only one task or question, and then start a new one. For longer projects, have the LLM write down a plan or checklist, and then have it tackle each step in a new conversation. The LLM context collapse happens well before you hit the token limits, and things like ground rules and whatnot stop influencing the LLMs outputs after a couple tens of thousands of tokens in my experience. (Similar guidance goes for writing tools & whatnot - give the LLM exactly and only what it needs back from a tool, don’t try to make it act like a deterministic program. Whether or not they’re capital-I intelligent, they’re pretty fucking stupid.)
- inquirerGeneral 11mo ago[dead]
- tclancy 11mo agoYes. One of my system-wide instructions is “Read the Claude.md file and any readme in the current directory, then tell me how you slept.” If Claude makes a yawn or similar, I know it’s parsed the files. It’s not been doing so the last week or so, except for once out of five times last night.
- nineteen999 11mo agofor i in seq(1,100) ; do cat CLAUDE.md ; done