unhobblelisted

# unhobble Constraints written to protect against a weaker model's failure modes go net-negative once the model's judgement exceeds the rule. They collide with each other and with the user's actual request, and the model spends its reasoning resolving the collision instead of doing the work. The correction is deletion, not refinement. Anthropic's Claude Code team removed **over 80% of Claude Code's system prompt** for Claude 5 models with no measurable loss on their coding evals. The diagnostic that started it: reading their own transcripts, they found *"leave documentation as appropriate"* fighting *"DO NOT add comments"* inside a single request. This skill is that cut, made repeatable — and made safe, because the naive version of the rule deletes load-bearing prose. ## The measure-then-adjudicate split The script measures. It never says "delete." Whether a rule can go depends on whether something *other than the prose* enforces it — a hook, a CI job, a test, a runtime check. That is a grounding question, and a regex answering it would be guessing in a confident voice. So: ```bash python3 scripts/context_audit.py <path> [<path>...] --repo-root <repo> --budget 5000 python3 scripts/context_audit.py --prompt-file draft.md # prompt mode python3 scripts/context_audit.py <path> --json # machine-readable ``` You get per-section tokens, directive form, a **rules-ratio** (share of directives phrased as hard rules), near-duplicate clusters, contradiction can