Marcos Paulo
ab2308441c
The repository had two skill directories. `skills/` held the four written for this project; `.agents/skills/` held the ones the agent context refers to. Nothing said which an agent should read, and `.agents/ORCHESTRATOR.md` only ever pointed at the second. Move the first four into `.agents/skills/` so there is one location, and add the vendored packages this chapter work used: `animation-vocabulary` and `improve-animations` (emilkowalski/skills), `teach` (mattpocock/skills), plus a local `translation` skill and `audit-translations.mjs` for the EN/PT pairs. `skills-lock.json` pins the vendored three by source and content hash, so a later re-vendor is a diff rather than a guess. `.claude/skills/` is symlinks into `.agents/skills/`, which is what makes them loadable here without a second copy on disk. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
1.8 KiB
1.8 KiB
name, description
| name | description |
|---|---|
| skill-reviewer | Review an Agent Skill package and produce a kind, evidence-backed improvement brief. Use when assessing a SKILL.md, its trigger, instructions, scripts, references, safety, or evaluation readiness; do not rewrite the package unless asked. |
Skill reviewer
Review the submitted package before proposing changes. Preserve the author's intent: this is a constructive assessment, not a replacement of their domain expertise.
Review flow
- Read
SKILL.mdand list bundled files. Check frontmatter validity, package-name alignment, and whether the description says both what the skill does and when it applies. - Identify the narrow job, the expected inputs, safe boundaries, a default workflow, and observable output. Mark any claim you cannot verify as a question, not a defect.
- Recommend only additions that change execution: a small RULES section for real invariants, a script for repeated fragile work, a reference for conditional detail, or eval cases for behavior that matters.
- Flag secrets, destructive actions, network calls, and unclear approval boundaries prominently. Never copy credentials into review artifacts.
- Return a friendly brief with: what already works, highest-value improvements, suggested package layout, and a small set of realistic test prompts.
Quality bar
- Prefer precise activation language over broad phrases such as "use for code."
- Keep the main instructions lean; send conditional or lengthy material to
references/and explain exactly when to read it. - Favor evidence and defaults over generic rules or tool menus.
- Recommend scripts only when they remove repeated, error-prone mechanics; document prerequisites and use relative paths.
Read the review rubric when scoring a package.