skill-builder

Invoke /skill-builder in CREATE mode. It interviews you about the problem, trigger, and failure modes, researches how top teams handle the domain, drafts the skill with STOP gates and anti-patterns, then validates it against the audit rubric before writing files.

Run the skill in AUDIT mode. It reads every SKILL.md in .claude/skills/, scores each across 9 dimensions using AUDIT-RUBRIC.md, classifies issues by priority, and presents a scorecard with recommended fixes before making changes.

Keep it under 250 characters with the primary use case in the first clause, 2-3 concrete deliverables, and an explicit trigger condition like 'Auto-triggers when...' or 'Invoke with /name'. The first 30 words must answer when to use the skill.

Procedural skills need STOP gates with explicit verifiable criteria before expensive, irreversible, or decision-changing steps. Referential and tool-reference skills should avoid rigid gates entirely and enforce behavior through anti-patterns and clear rules instead.

Advisory anti-patterns have a near-zero enforcement rate under time pressure. Rules using NEVER/ALWAYS/MUST with concrete consequences need a PreToolUse blocking hook, a ledger for auditing, and a PostToolUse recorder, as described in the three-piece enforcement pattern.

Split when SKILL.md exceeds 300 lines, when you have 3+ lookup tables or any table over 20 rows, or when 3+ historical lessons warrant a LESSONS.md. Every supporting file must be referenced by name in SKILL.md to avoid orphan files.