AIToday
AI Coding AssistantsZenn AI/MLPublished: Sep 30, 2026, 10:00 JST

CLAUDE.md grows 2 lines only for the 7 rules added

CLAUDE.md grows 2 lines only for the 7 rules added

3 Key Points

  1. What happened

    In the roughly two days after switching to a four-step placement test, the author added seven rules and pieces of knowledge — one landed in CLAUDE.md's rules, six went into documents the table points to, and CLAUDE.md grew two lines.

  2. Why it matters

    Under the four-step test, CLAUDE.md only grows at step 0 (top-priority) or step 1 (rules); step 2 adds nothing and step 3 adds a single table row, so the file stays short at the moment rules are added.

  3. What to watch

    The author notes only about two days have passed and says he cannot yet claim the file stays slim over a long period; the main text is the how-to-add-rules document opened before editing CLAUDE.md.

WHO IT HITSThis lands on developers and tech leads who maintain CLAUDE.md or similar AI agent instruction files, who now have a placement test to decide where each new rule goes instead of pruning a bloated file later.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Context & Analysis

The author's starting problem was not that CLAUDE.md was long, but that he wrote a reason along with every rule, so the file grew to about 17 times its initial length after roughly 20 rewrites over about 3.5 days. His first two rebuilds moved text around without changing it: splitting content into .claude/rules/ files left rules sitting beside explanations of why they exist, and a table of files to read per task ended up with rows that were not task names, including rows pointing at plan files that move away once done.

What he found is that the mechanical act of splitting does not reduce load. Files in .claude/rules/ without a paths: header are read at the start of every session, just like CLAUDE.md, and files pulled in with @ are inserted at launch; only Markdown-linked files stay closed until Claude needs them. That is why he ended up with three sections — top-priority rules, rules, and a table of files to read per task — and pushed all explanation into the documents the table points to. He also chose the table over paths:-tagged rules because a rule like running a checklist before push is not tied to opening a file, so a paths:-based rule might never be read.

The author himself flags that his evidence is thin: only about two days have passed since the current shape, and he states plainly that he cannot yet say it has stayed slim over a long period. The bet, it seems, is that deciding a rule's home at the moment of adding it costs less than sorting out a bloated file later — and that whether it holds will show in whether the two-line growth holds as the count of rules climbs.

FAQ
How much did CLAUDE.md grow after seven new items?
It grew two lines. One item went into the rules section and a short note was added before the table, while the other six went into referenced documents.
What belongs in the top-priority rules section?
Only rules that leak secrets, cannot be undone, or stop the whole mechanism — such as exposing API keys, writing to the production database without approval, or releasing without the checklist. Everything else is not promoted.

AI news that matters for your work, delivered every morning.

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

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.

Questions and answers are published on this page.

Related Articles

Next articleLessWrong post argues against strong decision-theoretic realism