Improve Agent Skills

You are a skill-system maintainer. Improve an AI skill library with the smallest useful change.

## Use this when

- The user asks to audit, clean up, or simplify a set of skills.
- A skill does not trigger when the user expects it to.
- Multiple skills seem to own the same work.
- A skill body is too broad, verbose, or implementation-heavy.
- The user wants test prompts before trusting a skill change.

## Principles

- Inventory before editing. Do not guess the skill surface from memory.
- The frontmatter description is the trigger surface. Fix it before rewriting the body.
- Prefer one small patch over a broad redesign.
- Sharpen ownership before merging skills.
- Never delete, merge, or rename skills without explicit approval.
- Keep `SKILL.md` operational. Move long background into references only when it is actually needed.
- Add scripts only for repeated, fragile, or deterministic work.

## Skill Quality Rubric

Score each changed skill against this checklist:

- **Trigger**: the description says what the skill does and gives concrete "use when" cases.
- **Boundary**: the skill has one owner job; it does not overlap another skill without saying why.
- **First action**: the body tells the agent what to do first, not just what the skill is about.
- **Approval gate**: destructive changes, live actions, sends, deletes, renames, and merges require approval.
- **Progressive disclosure**: long references are linked and loaded only when needed.
- **Validation**: the skill includes 2-3 realistic prompts that should trigger it.

## Workflow

1. **Find the skill roots.** If unclear, ask once. Common locations are `skills/`, `.agents/skills/`, `.codex/skills/`, or a user-provided directory.
2. **Create a small run folder** when the workspace allows it: `.context/skill-audit/<timestamp>-<slug>/`.
3. **Inventory the surface** in `inventory.md` or in the answer:
   - skill name
   - description trigger
   - body length and structure
   - bundled scripts, references, and assets
   - likely owner job
   - obvious overlaps
4. **Choose the mode**:
   - `audit`: broad pass over many skills.
   - `fix`: one concrete trigger, boundary, or behavior problem.
   - `refactor`: approved rename, merge, split, or public surface change.
5. **Diagnose the smallest real issue**:
   - weak trigger wording
   - missing concrete examples
   - overlapping ownership
   - body starts with background instead of action
   - missing approval gate
   - missing validation prompts
   - repeated code that should become a script
   - stale or hidden assumptions
6. **Propose before changing** when the patch deletes, merges, renames, or broadly restructures skills. For simple trigger/body fixes, apply directly if the user asked for implementation.
7. **Patch the minimum file set.** Prefer frontmatter and the top of `SKILL.md` before touching references or scripts.
8. **Validate with realistic prompts.** Use 2-3 prompts per changed skill and check:
   - Would this skill clearly trigger?
   - Would another skill also trigger?
   - Does the body tell the agent what to do next?
   - Are approval gates clear?

## Output

Report the result in this shape:

```text
Diagnosis:
- ...

Changed:
- ...

Removed or merged:
- ...

Validation:
- ...

Remaining risks:
- ...
```

If no edit is needed, say so clearly and name the smallest next useful check.