audit-skill

Skill Auditor

Audit a SKILL.md against agentskills.io best practices and produce structured findings ready for skill-creator iteration.

Process

1. Ask for the skill path if not provided
2. Read SKILL.md and any referenced files in the skill folder
3. Audit against each dimension below
4. Report findings, then ask: "Want to iterate on this with skill-creator?"

Audit Dimensions

1. Grounded in real expertise

• Instructions are specific: names actual APIs, schemas, edge cases, conventions
• No generic filler ("handle errors appropriately", "follow best practices")

2. Context window efficiency

• Adds what the agent lacks; omits what it already knows
• No over-explanation of common concepts (HTTP, PDFs, migrations, etc.)
• SKILL.md under 500 lines / 5 000 tokens
• Large reference material moved to references/ with conditional load triggers

3. Scope coherence

• Covers one coherent unit of work
• Not too narrow (would force multiple skills per task)
• Not too broad (hard to trigger precisely)

4. Instruction calibration

• Flexible where variation is acceptable — explains why rather than rigid steps
• Prescriptive only for fragile, sequential, or destructive operations
• One clear default per choice; alternatives mentioned briefly, not as equal menus

5. Procedures over declarations

• Teaches how to approach a class of problems, not what to produce for a specific instance
• Reusable method that generalizes across inputs

6. Effective patterns (check which are present and warranted)

• Gotchas — non-obvious environment-specific facts (wrong field names, soft deletes, misleading endpoints)
• Output templates — concrete template when format must be exact
• Checklists — step-by-step with checkboxes for workflows with dependencies or gates
• Validation loops — agent runs validator, fixes issues, repeats until passing
• Plan-validate-execute — for batch/destructive ops: structured plan → validation script → execution

7. Progressive disclosure

• Heavy content in separate files under references/ or assets/
• Each external file linked with a conditional trigger, not a generic "see references/"

Output Format

Report per dimension:

Dimension	Status	Finding	Fix
Grounded	OK / ⚠ / ✗	…	…

End with Top 3 priority changes ordered by impact.

Then ask the user whether to hand off to skill-creator for iteration.

Gotchas

• A skill that passes all dimensions may still not be valuable — if the agent already handles the task without it, the skill adds noise. Flag this if it seems likely.
• Don't penalize brevity. A 20-line skill can be excellent if it's specific and grounded.
• Progressive disclosure is only relevant for skills that legitimately need more content; don't suggest splitting a short skill.