Skill
Catch structural issues and format coherence problems in skill definitions before they cause downstream divergence
/skill-kit:skill-lint This skill reads a skill or prompt definition file (or directory), extracts all JSON schema blocks, and systematically checks every field against ten categories of issues spanning two phases. Phase 1 uses automated deterministic checks for frontmatter validation, step continuity, and cross-reference integrity. Phase 2 uses 5 parallel LLM subagents with model-matched complexity to detect MUST/MUST NOT contradictions, numeric fields quoted as strings, ungrounded fields, missing format constraints, inconsistent field names, open-ended enums, and structured data serialized as free-text. It produces a detailed report with severity ratings and concrete suggested fixes.
Phase 1 uses automated tools (Grep, Glob, regex) for definitive pass/fail structural checks. Phase 2 uses LLM judgment via parallel subagents for semantic analysis of JSON schema coherence. This separation ensures deterministic issues are caught reliably while nuanced problems get appropriate reasoning.
Semantic checks are distributed across 5 parallel subagents with models matched to check complexity: Sonnet for reasoning-heavy tasks (MUST consistency, cross-schema naming, verbatim-transfer semantics) and Haiku for pattern-matching tasks (numeric strings, open enums, structured data detection).
Covers frontmatter validation, step continuity, cross-reference integrity, MUST/MUST NOT consistency, numeric fields quoted as strings, ungrounded fields, missing format constraints, inconsistent field names, open-ended enums, and structured data serialized as free-text.
Every finding includes the exact current text, a clear explanation of the issue, and a concrete replacement suggestion. Fixes preserve semantic meaning while correcting structural and format coherence problems.
Accept a directory or file path and discover all relevant files.
The skill takes a path as argument, resolves whether it's a directory or file, and uses Glob to find all skill definition files. Explore subagents gather context about referenced tools and skills.
Run Phase 1 deterministic checks for frontmatter, step continuity, and cross-references.
Uses Grep, Glob, and regex to verify mechanically checkable properties. Checks frontmatter fields, step numbering sequences, and cross-reference integrity against known tool and skill names. Results are definitive pass/fail.
Launch 5 parallel subagents for Phase 2 LLM-driven checks.
Each subagent receives the file content and its specific category definitions. Models are matched to complexity: Sonnet for reasoning-heavy checks, Haiku for pattern-matching. All agents run in parallel for maximum throughput.
Merge all findings into a structured report with severity ratings and suggested fixes.
Results from both phases are combined into a single report with summary tables, per-finding detail blocks, and severity classifications. Categories with no findings are omitted for clarity.