ClaudeHack
0.000

Delete your CLAUDE.md, skills and hooks after a model upgrade, then add back only what breaks

Claude Code / CLAUDE.md·15 hours ago·verified 15 hours ago·easy·~15 min + a few days of observation setup·verified

Most CLAUDE.md lines exist to correct a model that no longer makes that mistake. Delete them after an upgrade and re-add only the ones a failure proves you need.

#claude-md#skills#hooks#context#prompting#ablation

The problem

CLAUDE.md, skills and hooks accumulate rules written to patch older models. They still load every session, costing tokens and boxing in a model that outgrew them.

The hack

Treat your setup as something to ablate, not to grow. After a model upgrade, delete CLAUDE.md, skills and hooks, work normally, and add a rule back only when Claude repeatedly stumbles on the same thing.

Why it works

Anthropic did exactly this to Claude Code itself: they cut roughly 80% of its system prompt for the newest model without losing coding-eval quality. Boris Cherny, who created Claude Code, recommends the same six-month cycle for user setups.

Setup

1. Commit your current setup so you can restore it. 2. Move `CLAUDE.md`, `.claude/skills/` and your hooks aside (keep hooks that enforce safety, like deny rules and destructive-command blocks). 3. Use Claude Code normally on real tasks for a few days. 4. Note every point where it stumbles. Ignore one-offs. 5. Re-add a rule only for mistakes that repeat, and keep the wording as narrow as possible. 6. Run `/claude-api prompt-audit` to flag remaining instructions written for older models and get the fixes as a diff.

Prompt

Re-read this CLAUDE.md line by line. For each line, ask whether you could have worked it out from the codebase yourself. List the lines you would delete.

Additional details

Boris Cherny's rule of thumb: every six months, delete your CLAUDE.md, delete your skills, delete your hooks, and see what the model does on its own. Two caveats worth keeping in mind: - **Ablate rules hardest, skills least.** `CLAUDE.md` loads on every session, so a stale line there costs you every time. Skills only load when they match, so bloat hurts less and can wait. - **Keep enforcement, delete instruction.** Hooks that block destructive commands and permission deny rules are policy, not advice for a model that will grow out of them. Those stay. If you want the automated version of the same idea, `/doctor` trims checked-in `CLAUDE.md` files by cutting content Claude could derive from the codebase, and migrates the rest into skills and nested `CLAUDE.md` files that load on demand.

Source

youtube·View original →·by Boris Cherny (Y Combinator)

Does this still work?

0 said it works · 0 said it's broken

Related

Discussion

No comments yet. Be the first to say whether this worked for you.