You wrote "never push without a review" into CLAUDE.md. Claude pushed. So you added a line, then another.
The file is 380 lines now and the model follows less of it than it did at 40.
CLAUDE.md is context, not configuration.
Anthropic's docs say it in those words. It arrives as a message after the system prompt, with no guarantee of strict compliance.
The file is not broken. You have been asking it to do a job it was never built for. That's why the file gets skipped.
In this issue
→ 1. The one-third rule
→ 2. Three homes for every instruction
→ 3. The audit, run on my own setup
→ 4. The five-minute audit
→ The four ways a file goes quiet
Four moves, in order
1. The sizing rule, one-third
On September 2, Anthropic published a guide to building agents. One paragraph is about where instructions live.
Their rule: "anything relevant to a third or more of your traffic, whether anticipated before launch or observed in production, goes in the system prompt, and the rest goes in skills."
If a line matters in a third of your sessions or more, it stays. If it matters in fewer, it moves.
↳ Trap: the docs target under 200 lines, and every imported file counts.
2. Three homes for every instruction
CLAUDE.md holds what you would otherwise re-explain. Build commands, conventions, layout, the always-do rules. Each line checkable.
Skills hold procedures. One file each, loaded only when the job matches.
Hooks and deny rules hold what must happen. "Never push to main" carries force the model cannot opt out of. A PreToolUse hook or a deny rule is enforced regardless of what Claude decides.
↳ Trap: if you are capitalizing NEVER in a markdown file, the word is telling you it belongs in a hook.
Same file, different name. Your system prompt, your custom instructions, your .cursorrules are the always-on file. A skill or a pasted doc is the procedure.
Where the tool has no hook layer, the third home is a script, or a checklist you run yourself.
3. The audit, run on my own setup
The fix, ruled August 21: about ten gates in code, everything else demoted to data one step reads when it needs it.
The sort, three piles.
- To code: a brief check before any post renders, the ship gate, the render cap, this newsletter's own gate.
- To data: the banned-word list and the design rules, read only when a step needs them.
- Marked for deletion: 40 self-audit triggers and roughly 200 prose rules.
Two gates bit back.
Receipt: one false-blocked 19 of 21 proven winners.
One rewrote a hook that had done 38,000 impressions, and the rewrite was weaker.
Data outranks the rule.
Receipt: root file 347 lines, nine rules files beside it, 962 lines loaded at launch.
Thirty lines still contain never or always.
↳ Run the same three numbers on yours.
4. Do this now: the five-minute audit
☐ wc -l CLAUDE.md. Over 200 and you already know why the rules slip.
☐ Run /context. Every file you expected under Memory files is there, or it never loaded.
☐ Each line: does it apply in a third of my sessions? No: skill or path-scoped rule.
☐ Each never and always: does skipping it cost money or a repo? Yes: hook or deny rule.
☐ Same topic twice? Keep the newer line, delete the older.
The four ways a file goes quiet
- It grew past 200 lines. Adherence drops with length.
- Two lines contradict. Claude may pick one arbitrarily.
- It is vague. "Format code properly" cannot be verified. "Use 2-space indentation" can.
- It never loaded. Subdirectory files load only when Claude reads there.
The file that survives the audit is short, and the model follows it.
The writing did not improve.
There is less to ignore.
Reply and tell me the one rule your AI keeps skipping. I read every reply and I will tell you which home it belongs in.
The full edition carries the paste-ready skeleton, the audit with today's numbers, and where this falls short.