| name | optimizing-skill-repo |
| description | Audits and optimizes a collection of agent skills for compliance, concision, and cross-tool portability. Applies Agent Skills spec, Anthropic best practices, and progressive disclosure principles. Use when reorganizing skills, auditing skill quality, or optimizing a skill catalog. |
Optimize Skill Repo
Systematically audit and improve a skill collection. Interview the user at each decision point — provide recommended answer, wait for confirmation before proceeding.
Process
Optimization checklist:
- [ ] Phase 1: Structure — flatten, remove scaffolding, decide repo purpose
- [ ] Phase 2: Inventory — classify each skill (keep/merge/drop/rewrite)
- [ ] Phase 3: Research — fetch latest spec + best practices
- [ ] Phase 4: Audit — check every skill against principles
- [ ] Phase 5: Rewrite — apply fixes, deduplicate, add examples/checklists
- [ ] Phase 6: Verify — final compliance pass
Phase 1: Structure
Resolve these decisions with the user:
- Repo purpose — browsable catalog, drop-in kit, or package source?
- Folder layout — flat at root (recommended for catalogs) or grouped?
- Non-skill files — which docs, configs, wrappers to keep/drop/rewrite?
- Single README — minimal, no per-skill descriptions (catalog grows)
Phase 2: Inventory
For each skill, classify:
- Keep — distinct purpose, non-obvious process, earns its tokens
- Merge — overlaps significantly with another skill
- Drop — redundant, project-specific, or teaches what models already know
- Rewrite — good concept but needs generalization or trimming
Present classification table. User approves before proceeding.
Phase 3: Research
Fetch latest standards:
- Agent Skills spec at agentskills.io/specification
- Anthropic best practices at platform.claude.com skill authoring docs
- Any other authoritative agent instruction guidelines
Extract actionable principles (see Audit Principles below).
Phase 4: Audit
Check every skill against:
Spec compliance:
name matches directory, lowercase + hyphens, ≤64 chars
description ≤1024 chars, third person, "Use when [triggers]"
- SKILL.md body under 100 lines (500 per spec, 100 for tight catalogs)
- References one level deep from SKILL.md
Content quality:
- No teaching known things (SOLID, DRY, Nielsen's heuristics — models know these)
- Imperative language ("Analyze X" not "You should analyze X")
- Consistent terminology throughout
- No time-sensitive information
Effectiveness patterns:
- Workflow checklist for multi-step processes
- Input/output examples for skills where output format matters
- Match freedom to fragility (high freedom for judgment, low for fragile ops)
- Progressive disclosure (details in REFERENCE.md, not SKILL.md)
Cross-skill health:
- No content duplicated across skills
- Each skill has distinct, non-overlapping purpose
- Descriptions differentiate clearly for agent discovery
Present audit table with issues per skill.
Phase 5: Rewrite
Apply fixes in batch:
- Rename folders to gerund form (Anthropic convention)
- Strip known-to-models content
- Deduplicate shared content (pick one owner skill)
- Add checklists and examples where missing
- Split overlong SKILL.md into SKILL.md + REFERENCE.md
- Generalize project-specific skills or drop them
Phase 6: Verify
Final pass:
- Line counts all within limits
- Name-directory consistency
- Description compliance
- No orphaned references
- Structure is flat and clean