Rule Curator

A skill that audits the behavioral rules an agent operates under and helps you keep, drop, or rewrite them.

Agents accumulate rules across many files: global and project CLAUDE.md, a memory system, project docs like AGENTS.md, ADRs, conventions. Nobody prunes them. Rules go stale, contradict each other, duplicate installed skills, or restate things the model already does by default. Every rule also costs context tokens each session, and the more rules there are, the less attention each one gets.

Rule Curator gathers the rules, analyzes each one against the whole set, and builds a page in your browser where you mark each rule keep, drop, or modify. It then turns your decisions into a reviewable edit plan and, once you confirm, applies the edits to the source files.

What it does

Every rule gets two separate write-ups:

1. A behavioral impact statement: what following the rule actually changes about the agent's behavior, and what it costs. Every rule gets one.
2. An audit note, only when something is wrong, from a fixed set of categories: redundant, stale or superseded, misfiled, low-yield ritual, not actionable, brittle, or conflicting.

The skill checks the categories that audits usually skip (brittle rules, misfiled entries), and it does not invent problems on healthy rules to look thorough.

If your rule files live in git, the audit records the commit it ran against (meta.commits in rules.json, shown in the UI header and included in both exports). When you come back to your decisions days later, that record shows exactly which files changed in the meantime. If your rule files are not in git, the skill says so up front and warns you before applying anything that a deleted rule cannot be recovered.

Repo contents

• SKILL.md — the workflow, the analysis contract, and the disciplines
• references/analysis-contract.md — the full rubric with worked examples
• references/apply-plan-format.md — the edit-plan format and per-verdict rules
• build_curator.py — standard-library script that builds the self-contained UI
• curator-template.html — the UI template (data is injected at build time)
• evals/ — a synthetic test corpus with planted issues, its answer key, and recorded test runs; evals/make_history.py rebuilds the corpus with a reproducible git history for the version-history test cases (not yet run)
• docs/specs/ — the design spec

Install

Clone into your Claude Code skills directory:

git clone https://github.com/bwdurst/rule-curator.git ~/.claude/skills/rule-curator

or copy the folder there manually. Then ask Claude to audit, prune, or clean up your rules, or run /rule-curator.

One thing worth doing first: if your rule files are not under version control, run git init in the directory that holds them (your memory directory, for example). With history, the audit can pin exact commits, detect edits made while you were deciding, and revert a bad prune. Without it, a deleted rule is gone for good.

Build the curation UI directly

You can also run the build step yourself once you have a rules.json:

python build_curator.py rules.json -o rule-curator.html

This injects the data into curator-template.html and writes one self-contained file you can open offline. It uses the standard library only, so there is nothing to install.

How it was built

This skill was developed test-first. The evals/ directory holds a synthetic rule corpus with planted issues across every audit category, an answer key, and the recorded baseline (without the skill) and verification (with the skill) runs on both a strong and a weak model. See evals/RED-baseline-findings.md and evals/GREEN-verification-findings.md.