spellbook: Skill for Claude Code

Audit Spellbook

You are auditing the spellbook project itself. This skill orchestrates parallel subagents to comprehensively analyze skills, commands, docs, and prompts for optimization opportunities.

Invariant Principles

1. Parallelism maximizes audit coverage - All audit agents launch simultaneously; sequential execution wastes context
2. Token efficiency compounds - Small savings multiply across always-loaded descriptions, skill bodies, and runtime
3. CSO prevents workflow leak - Descriptions trigger only; workflow in description = Claude follows description not skill
4. Evidence over claims - Every finding requires file/line/example proof; no unsubstantiated optimization recommendations
5. Actionable over diagnostic - Report must produce implementable items with clear priority

Trigger Conditions

Use this skill when:

• User asks to "audit spellbook", "optimize skills", "review spellbook"
• Before major releases to ensure quality
• When concerned about token usage or instruction bloat
• Periodically for maintenance

Execution Flow

Phase 1: Launch Parallel Audit Subagents

Launch ALL of these subagents in a SINGLE message (parallel execution):

1. Factcheck Agent

Audit all documentation in spellbook for factual accuracy.

Files to check:
- README.md
- docs/**/*.md
- CHANGELOG.md
- Any claims in skill/command descriptions

For each claim found:
1. Identify the assertion
2. Verify against: code, external sources, logical consistency
3. Flag unverifiable or incorrect claims

Output: JSON array of {file, line, claim, status: "verified"|"unverified"|"incorrect", evidence}

2. Instruction Engineering Compliance Agent

Audit all instruction files against instruction-engineering principles.

Files: skills/*/SKILL.md, commands/*.md, rules/*.md

Check for:
- Clear role definition
- Explicit trigger conditions
- Structured output formats
- Edge case handling
- Appropriate use of examples (not excessive)
- Action-oriented language
- Avoidance of ambiguity

Output: JSON array of {file, issues: [{principle, violation, suggestion}], score: 0-100}