Skill Maintenance — CowboyLogic/ai-dev

Run the full maintenance cycle for one skill in this repository. The skill to maintain is named by the argument this skill was invoked with. This file calls that name <skill-name>, and the skill lives at skills/<skill-name>/. If no argument was given, ask which folder under skills/ to maintain before doing anything else.

Work through each stage in order. Do not skip a stage or move to the next until the current one is complete.

Stage 1 — Research and update the skill

Read skills/<skill-name>/SKILL.md and all files under skills/<skill-name>/references/ in full to understand the current state.
Research what has changed since the skill was last updated. The skill's sources.json (if present) lists the upstream docs to check. Use the Tank research subagent to pull current information from those sources. Focus on: new config fields, changed defaults, deprecated keys, new CLI commands, and anything that would cause the current reference files to produce wrong or outdated answers.
Apply updates:
- Edit skills/<skill-name>/SKILL.md if the task→reference map, quick-ops list, or config file table needs updating.
- Edit files under skills/<skill-name>/references/ with specific, targeted changes — do not rewrite sections that are still accurate.
Note every change made — you will need this list for the changelog entry in Stage 4 and the commit message in Stage 5.

Stage 2 — Run benchmark evaluations

Use the eval cases in skills/<skill-name>/evals/evals.json. If no evals exist yet, draft 3–4 realistic test cases that cover the skill's primary tasks and save them to skills/<skill-name>/evals/evals.json before proceeding.

All runs go in skills/<skill-name>-workspace/iteration-N/ where N is the next iteration number (check for existing iteration folders first).

Launch all runs in a single turn — spawn with-skill AND without-skill subagents simultaneously for every eval case. Do not batch them sequentially.

With-skill run: provide the full SKILL.md content and relevant reference file content in the subagent prompt, then ask it to answer the eval prompt.
Without-skill (baseline) run: same prompt, no skill content — the subagent answers from training knowledge alone.

While runs are in progress, draft assertions for each eval case and save them to evals/evals.json. Good assertions are:

Objectively verifiable (not "does it sound right")
Named clearly enough that someone reading the benchmark immediately knows what each one checks

Write eval_metadata.json in each eval directory with the prompt and assertions.

Stage 3 — Grade and benchmark

Once all runs are complete:

Grade every run against its assertions. For each run directory write grading.json with this structure:

{
  "eval_id": "<name>",
  "variant": "with_skill | without_skill",
  "assertions": [
    { "id": "<id>", "passed": true|false, "note": "<brief evidence>" }
  ],
  "score": N,
  "max_score": N,
  "pass_rate": 0.0–1.0
}

Write benchmark.json in the iteration folder:

{
  "skill_name": "<skill-name>",
  "iteration": N,
  "run_date": "<YYYY-MM-DD>",
  "evals": [
    {
      "id": "<eval-name>",
      "with_skill":    { "score": N, "max_score": N, "pass_rate": 0.0 },
      "without_skill": { "score": N, "max_score": N, "pass_rate": 0.0 },
      "delta": 0.0
    }
  ],
  "summary": {
    "with_skill_total": N,
    "without_skill_total": N,
    "max_total": N,
    "with_skill_pass_rate": 0.0,
    "without_skill_pass_rate": 0.0,
    "overall_delta": 0.0
  }
}

Summarise the results: overall pass rates, delta, and — importantly — what the baseline got wrong and why it matters.

Stage 4 — Record history and results, and update the docs overview page

The skill folder is the record of the skill's history and benchmark results. The docs overview page is short and links to it.

In the skill folder

skills/<skill-name>/CHANGELOG.md — add an entry at the top, newest first, listing every change made in Stage 1. Create the file with a # Changelog heading if it does not exist.
```
## <YYYY-MM-DD> — v<N> (<label>)

- <File changed and what was updated.>
```
skills/<skill-name>/README.md — add the Stage 3 results under an ## Evaluation results heading: the run date and iteration, the scenario and run counts, the overall pass rates, a per-scenario table with notes on what the baseline got wrong, and a key takeaway. Keep earlier iterations below the new one. Create the file if it does not exist.

The docs overview page

The page lives at docs/skills/<skill-name>.md. Check whether it already exists. It lets a reader decide whether to use the skill and install it. It must not embed skill content: no instruction text, reference file contents, examples, workflows, rule tables, or script usage. Link to skills/<skill-name>/ instead.

If it does not exist — create it

Use this structure:

# <Human-readable skill title>

<One-paragraph description of what the skill covers.>

- **Skill name:** `<skill-name>`
- **Source:** [skills/<skill-name>](https://github.com/CowboyLogic/ai-dev/tree/main/skills/<skill-name>)

---

## What it does

<One short paragraph: the gap the skill closes and how it works. List the topics it
covers. Do not reproduce rules, tables, or examples from the skill.>

## Where it applies

<One or two sentences: the clients, tasks, or repositories the skill is for.>

---

## Install

### Using `npx skills` (recommended — works across all agents)

\`\`\`bash
npx skills add CowboyLogic/ai-dev --skill <skill-name> -g
npx skills add CowboyLogic/ai-dev --skill <skill-name> --agent <agent> -g
npx skills add CowboyLogic/ai-dev --skill <skill-name> -g -l
\`\`\`

### Verify installation

\`\`\`bash
npx skills ls -g
\`\`\`

---

## Evaluation and history

<Include this section only if the skill has benchmark results. One sentence giving the run
date and the with-skill and baseline pass rates, then links to the skill's README.md and
CHANGELOG.md on GitHub.>

---

## Related

<Links to sibling skills, if any.>

Follow all repo Markdown standards:

GFM callouts (> [!NOTE], > [!TIP], > [!WARNING]) — never plain bold blockquotes for informational content
Blank lines around all block elements
Fenced code blocks with language identifiers

If it already exists — update it

Do not add a "Last updated" date. CHANGELOG.md and git history carry dates.
Update the one-line result in "Evaluation and history" with the new benchmark data from Stage 3.
Update "What it does" and "Where it applies" if the skill's scope changed materially.

Wire up the page (new pages only)

After creating a new page:

Add it to the Skills: section in mkdocs.yml:

- Skills:
    - Overview: skills/index.md
    - <Title>: skills/<skill-name>.md   # add here

Update the skill's entry in docs/skills/index.md:
- Change the existing GitHub-only link to include a Skill Overview link:
```
[Skill Overview](skills/<skill-name>.md) · [View on GitHub](https://github.com/...)
```
- If there is no entry yet, add one in the appropriate section.

Stage 5 — Propose a commit message

Review all files changed across all stages. Propose a commit message following the repo convention:

docs(skills): <imperative subject line, ≤50 chars>

<Body — what changed and why, bullet points per file/area>

Use type docs(skills) when the primary changes are docs + skill reference updates. Use feat(skills) only if a new skill was created from scratch.

Present the message for review. Do not run git commit — wait for explicit confirmation from the user.

skill-maintenance

SKILL.md.github/skills/skill-maintenance/SKILL.md