The failure mode is asymmetric accretion. Every bad output earns a new rule and nobody ever removes one, so months later the model is trying to satisfy dozens of instructions on every response whether they apply or not.
Test the trim as a batch, not rule by rule: delete everything flagged at once, run your three most common tasks, and add back only the one rule whose output actually got worse. Removing one at a time hides the effect.
Model upgrades quietly obsolete instructions, so workarounds written for last year's weaknesses keep getting enforced. The right trajectory for an instruction file is to shrink over time, not grow.
Why it matters
A self-audit prompt that has Claude find and cut the redundant, conflicting instructions you have accumulated in your setup. Run it if your CLAUDE.md has quietly grown into a contradictory mess — most have.
Key quotes
“I deleted half my Claude setup last week and every output got better.”
“It's like handing a chef a 47-step recipe when they only need 12.”
“Anthropic's own engineering team recently found the same thing in their internal Claude setups.”
“The people getting the best outputs are the ones who regularly cut the dead weight.”
