Audit Checklist

Trigger contract

• Frontmatter description states when to use the skill, not the workflow.
• Name is short, hyphen-case, action-oriented, and discoverable by likely user words.
• Body does not rely on a "when to use" section to fix weak metadata.

Progressive disclosure

• SKILL.md holds only the main workflow, routing rules, and required invariants.
• Detailed variants, schemas, examples, style rules, and long checklists live in directly linked references.
• References are one hop from SKILL.md, named by the decision that loads them, and non-empty.

Workflow and delegation

• The main workflow is explicit enough that an agent knows what to do first, what gates the next phase, and when to stop.
• Subagents are part of the workflow only when they reduce real scope risk; they are not a decorative list of roles.
• The lead agent owns final consistency, evidence quality, and output claims.

Template ownership

• Base report templates own structure and parser-safe fields.
• Runtime-specific templates own runtime rules such as Canvas, HTML, SDK imports, companion bullets, preview, and validation.
• Style templates own visual grammar only. They should not duplicate runnable skeletons, shared SDK rules, or mode-selection logic.

Prompt and report readability

• Cut explanations an AI can infer from headings, examples, or local code.
• Prefer one concrete contract over repeated warnings in several files.
• Replace placeholder walls, consecutive empty headings, and giant table rows with compact field requirements or one realistic example.
• Keep exact labels, scoring scales, and parser-sensitive metadata centralized.

Evidence and validation

• Every important review claim has a file path, line, command, diff, or artifact anchor.
• Missing evidence is reported as missing or unverified, not filled by inference.
• Validation checks the same surface the skill claims to support. For visual chains, build/preview/runtime validation is separate from Markdown review.