+1 we ended up adding something to our CLAUDE.md specifically to counter it. Something along the lines of "comments should be treated the same as production code, they shouldn't leak the content of conversations and stand the test of time". It doesn't always work, as other people say claude stops following these the more its context window gets filled, but now it's mostly a matter of telling it in a new session to review the changes against the guidance on comments and clean them up.

We have it as well, but it almost never get respected and we still see those lengthy comments. Anyone seen better luck?

I mean I do the same at work where we currently use Anthropic, but we are also on a token optimisation drive so it’s getting hard to justify spending even more tokens cleaning up after Anthropic’s garbage by paying Anthropic more. We have started having users use GPT 5.6 Sol to compare quality and effficiency over Opus. Fable not happening at work because of the data sharing clauses and the completely nuts pricing.

I find it particularly bad at writing test comments, I usually start fixing bugs by adding regression tests and Opus tend to put story into comments for these, including “but it fails because…” followed by bug description including line numbers and call chains. It just lives to narrate everything

Same, oh God the leaking of our chat infuriates me. It also loves to document public functions with what the caller is and doing, and I'm like that's not how this works! The point of an abstraction or interface is that it stands on its own. It's not supposed to be for some specific caller elsewhere in the code.

This, idk how many times I've begged it to let the code be self documenting. The worst? claude's own comments cause it to go in unproductive loops. It tries to satisfy for your instructions and the incorrect comments it assumes are yours but forgot it added them.

The key phrases I've found that help is to tell it "No memorializing decisions in output artifacts" and "Only describe what is, never what was in documentation"

Comments which, of course, only further increase token consumption when again processed ...

Whenever I have Claude make changes I ask Codex to review the changes and reduce the comments.