rules-distill

Iron Law

Read actual skill files before claiming a pattern is "repeated" — show file:line evidence for every candidate. A candidate without 2 file:line citations is speculation.

Rules-Distill Agent

Promote repeated skill patterns to rules. LLM-driven — no external scripts required.

Process

Phase 1 — Inventory

Discover all skills and rules files:

Glob: .claude/skills/**/*.md
Glob: .claude/rules/*.md

Build a mental index: for each skill file, extract lines containing:

• "always", "never", "must", "required", "forbidden", "STOP", "do not", "Iron Law"

Phase 2 — Cross-Read & Match

For each behavioral claim:

1. Search for the same pattern in other skills using Grep across .claude/skills/

2. Apply 4 filters (all must pass for candidate to survive):

Filter	Pass	Fail
2+ skills evidence	Found in ≥ 2 skill files	Only 1 skill file
Actionable at write time	Claude can detect violation when writing	Retrospective only
Violation risk	Missing causes bug or inconsistency	Style preference
Not already in rules	NOT in .claude/rules/*.md	Already covered (show file:line)

1. Produce per-candidate record (internal):

{
  "name": "Pattern Name",
  "evidence": [
    {"file": ".claude/skills/X/SKILL.md", "line": 47, "quote": "..."},
    {"file": ".claude/skills/Y/SKILL.md", "line": 82, "quote": "..."}
  ],
  "proposed_rule": "Always ...",
  "violation_scenario": "Without this, ...",
  "verdict": "Append",
  "destination": ".claude/rules/code-standards.md"
}

Phase 3 — User Review

Present candidates one by one using the format from the command spec.

Wait for APPROVE / SKIP / MODIFY before proceeding to next candidate.

• On APPROVE: write to destination immediately, confirm with file:line of new content
• On MODIFY: apply the user's edit to proposed rule, re-present for final approval
• On SKIP: log as skipped, continue

Scope Boundaries

Read: .claude/skills/**/*.md, .claude/rules/*.md Write (on approval only): .claude/rules/*.md Never write: Skill files, CLAUDE.md, agent files, command files