Skill: Refine Skill Design

Purpose

Audit and improve an existing SKILL.md to maximize its ASQM score. Apply a structured meta-audit across 4 dimensions: intent clarity, behavioral logic, constraint precision, and example quality. Produce an optimized version with a clear diff summary.

Core Objective

Primary Goal: Take any SKILL.md and return an improved version scoring >=17 ASQM (validated threshold).

Success Criteria:

1. Every section is present and well-formed (frontmatter, Purpose, Triggers, Behavior, Restrictions, Self-Check)
2. Intent is unambiguous — another agent reading this skill would behave identically
3. Constraints use hard/soft boundary separation
4. At least 2 examples including 1 edge case
5. The diff summary explains every change with rationale

Acceptance Test: The refined skill passes verify-skill-structure.mjs and scores >=17 on ASQM dimensions.

Triggers

• Called after format-skill generates a new SKILL.md (quality gate before PR)
• Called manually when a skill's ASQM score drops or needs improvement
• Called by curate-skills when a skill fails quality checks

Behavior

Step 1: Parse and validate structure

Check the SKILL.md for required sections:

✅ YAML frontmatter (name, description — Agent Skills standard; license, metadata optional)
✅ Purpose section
✅ Triggers section
✅ Behavior section (with numbered steps)
✅ Restrictions section (hard/soft boundaries)
✅ Self-Check section (checkbox list)
⚠️ Examples section (at least 2, including 1 edge case)

Step 2: Meta-audit across 4 dimensions

1. Agent-native (0-5): Can an AI agent execute this skill without ambiguity?

• Are triggers machine-detectable?
• Are actions concrete (not vague like "consider" or "think about")?
• Are outputs well-defined?

2. Cognitive (0-5): Is the skill cognitively well-structured?

• Single responsibility (one clear job)?
• Appropriate abstraction level (not too broad, not too narrow)?
• Clear decision points (if X then Y, not "use judgment")?

3. Composability (0-5): Does it play well with other skills?

• Clear input/output contract?
• Named handoff points to related skills?
• No overlap with existing skills (check related_skills)?

4. Stance (0-5): Are constraints precise and enforceable?

• Hard boundaries use "NEVER", "MUST", "ALWAYS"
• Soft boundaries use "prefer", "when possible", "by default"
• No contradictions between constraints

Step 3: Apply improvements

For each dimension scoring <4, make targeted improvements:

• Rewrite vague triggers as concrete conditions
• Replace advisory language with actionable steps
• Add missing examples or edge cases
• Separate hard from soft constraints
• Add self-check items for gaps

Step 4: Output

Present the result as:

## Refine Report: [skill-name]

### Scores (before → after)
| Dimension     | Before | After | Change |
|---------------|--------|-------|--------|
| Agent-native  |   3    |   5   | +2     |
| Cognitive     |   4    |   4   |  0     |
| Composability |   3    |   5   | +2     |
| Stance        |   4    |   5   | +1     |
| **Total**     | **14** | **19**| **+5** |

### Changes
| Section | Change | Reason |
|---------|--------|--------|
| Triggers | Added concrete conditions | Was too vague for agent detection |
| Restrictions | Split hard/soft | Mixed severity confused boundaries |
| Examples | Added edge case | Only had happy path |

### Version Suggestion
1.0.0 → 1.1.0 (minor: improved quality, no behavioral change)

Restrictions

Hard boundaries

• NEVER remove existing examples or restrictions without replacement
• NEVER change the core purpose/intent of the skill
• NEVER introduce dependencies on external services
• NEVER auto-apply changes — always present diff for human review

Soft boundaries

• Prefer minimal changes over rewrites (smallest diff that maximizes score)
• Prefer adding examples over rewriting existing ones
• When two phrasings score equally, keep the original

Self-Check

• [ ] All required sections present and well-formed?
• [ ] Each ASQM dimension scored and justified?
• [ ] Diff summary includes rationale for every change?
• [ ] Version bump suggestion follows semver?
• [ ] Refined skill passes structural validation?
• [ ] No meaning/intent drift from original?

Examples

Example 1: Improving a weak trigger

Before (agent_native: 2):

## Triggers
Activate when you notice a deployment issue.

After (agent_native: 5):

## Triggers
Activate when ALL of the following are true:
- A `pnpm build` or `npm run build` command was just executed in `plugins/`
- The build output does NOT include the expected extension files in `extensions/`
- OR: a deployment command was run without a prior build step in the session

Example 2 (edge case): Skill that's already good

Input: A skill scoring 18/20 (validated).

Output:

## Refine Report: preserve-directus-code-identifiers

### Scores (before → after)
All dimensions >=4. No changes needed.

### Recommendation
This skill is already at validated quality. Consider:
- Adding a 3rd example if the skill is frequently referenced
- No version bump needed