alt Hacker News> 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.
Replies
Yeah, 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.
Yes. 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.
The 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!”
The 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?
One 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.)