| name | rules-distill |
| description | Scan skills to extract cross-cutting principles and distill them into rules -- append, revise, or create new rule files. |
| origin | ECC |
Rules Distill
Scan installed skills, extract cross-cutting principles that appear in multiple skills, and distill them into rules.
When to Use
- Periodic rules maintenance (monthly or after installing new skills)
- After a skill-stocktake reveals patterns that should be rules
- When rules feel incomplete relative to the skills being used
How It Works
Phase 1: Inventory (Deterministic Collection)
Scan all skills and rules, count files and headings.
Phase 2: Cross-read, Match and Verdict (LLM Judgment)
Group skills into thematic clusters. For each cluster, extract principles that:
- Appear in 2+ skills
- Are actionable ("do X" or "don't do Y")
- Have clear violation risk
- Are not already in rules
Assign verdicts: Append, Revise, New Section, New File, Already Covered, or Too Specific.
Phase 3: User Review and Execution
Present a summary table. User approves, modifies, or skips each candidate. Never modify rules automatically.
Verdict Types
| Verdict | Meaning |
|---|
| Append | Add to existing section of existing rule file |
| Revise | Fix inaccurate/insufficient content |
| New Section | Add new section to existing rule file |
| New File | Create new rule file |
| Already Covered | Sufficiently covered in existing rules |
| Too Specific | Should remain at the skill level |
Design Principles
- What, not How: Extract principles only; code examples stay in skills
- Link back: Draft text includes
See skill: [name] references
- Deterministic collection, LLM judgment: Scripts guarantee exhaustiveness; LLM guarantees contextual understanding
- Anti-abstraction safeguard: 3-layer filter prevents overly abstract principles
