
What happened
The author's CLAUDE.md file for Claude Code grew from 4,653 characters on 2026-08-14 to 50,702 characters on September 18, and its warning marks went from 1 to 203.
Why it matters
The writer says a memo where everything is marked important makes it impossible to tell what matters, so each addition may have diluted the rest—meaning more rules may not mean better compliance.
Summaries like this, in your inbox every morning.
The file started as a handover note: Claude Code reads CLAUDE.md in full at the start of every conversation, so it is where a user writes "always do this" instructions. The author added a section each time something went wrong—why the rule was needed, when it happened, what not to do—and each entry looked indispensable on its own, because removing it risked repeating the same failure.
That incremental, locally rational habit produced the growth: 4,653 characters on 2026-08-14 became 50,702 characters on September 18, and the "⚠️" marks used to flag important points went from 1 to 203. The author's own reading is that a memo where everything is flagged leaves no way to tell what is important, so each addition may have weakened the rest.
The attempted fix was to separate rules that only apply "when X" into separate procedure files read only in those situations. But the growth did not stop: the file was 26,419 characters when that decision was made and 38,527 characters two days later. The details of how 51,598 characters were eventually cut to 7,016—and whether it has stayed down—are covered in a follow-up piece.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Deepseek shipped Deepseek-V4.1-Flash-1017 via its website, app, and API, and published the weights on Hugging…

On John Collison's Cheeky Pint podcast, Zuckerberg and Wang said a 44-minute call and a memo led to Meta Super…

The developer switched ALPHA FORGE's judge model to the brand-new Opus 5.5 on Saturday, Sept 26, and it return…

Across 100 chances to answer questions absent from a six-line sample document, Claude Code on claude -p versio…

An AI running on Claude Code described its parts: a Codex watchdog checking every 1 minute, ChatGPT as design…

In the developer's test on Claude Code 2.1.302, deny rules stopped 9 of 11 risky commands, but /bin/rm secrets…
