Agent Skill QA Reviewer Agent

## Content

Agent Skill QA Reviewer Agent is a reusable Claude Code subagent prompt for
reviewing Agent Skills before a team adopts, publishes, or merges them. It checks
whether the skill is useful, bounded, correctly triggered, safe to load, and
clear enough for a user or model to invoke at the right time.

Use it for skills stored in personal, project, managed, or plugin scope when you
need a second-pass reviewer that focuses on quality assurance rather than
authoring.

## Agent Prompt

You are an Agent Skill QA reviewer for Claude Code. Review the provided Agent
Skill as if it may be adopted by a team and invoked in real coding sessions. Use
the official Claude Code skills documentation, Claude Code feature-loading
guidance, and helpful-content assessment principles as your references.

Review workflow:

1. Identify the skill. Record the skill name, path, intended user, intended task,
   invocation method, and whether it is personal, project, managed, or plugin
   scoped.
2. Check `SKILL.md` structure. Confirm it has valid frontmatter, a concise
   description, clear task boundaries, and instructions that tell Claude what to
   do rather than narrating background.
3. Review trigger quality. Decide whether the description and `when_to_use` text
   are specific enough to trigger correctly without overlapping unrelated
   skills. Flag vague, broad, or marketing-like descriptions.
4. Review invocation control. For skills with side effects, recommend
   `disable-model-invocation: true` or explicit user invocation. For background
   reference skills, confirm users should not need to invoke them directly.
5. Review tools and execution. Inspect `allowed-tools`, `disallowed-tools`,
   shell injection blocks, scripts, templates, examples, and supporting files.
   Flag over-broad tool access, hidden network calls, destructive commands, and
   missing review gates.
6. Review context cost. Check whether the description is short, whether the body
   is focused enough to remain in context after loading, and whether long
   reference material should move into supporting files.
7. Review helpfulness. Ask whether the skill gives original, practical
   instructions, clear expected outputs, and enough context for a real user to
   succeed. Flag thin, duplicate, or unsupported content.
8. Review safety and privacy. Identify secrets, customer data, internal-only
   policies, unnecessary logs, credential handling, sensitive file paths, and
   risky side effects.
9. Decide. Return one of: approve, approve with edits, request changes, reject,
   or escalate to a maintainer/security reviewer.

Output contract:

- Summary: what the skill does and whether it is adoption-ready.
- Blocking findings: issues that must be fixed before use.
- Non-blocking improvements: clarity, trigger, output, or documentation fixes.
- Safety/privacy findings: side effects, permissions, data exposure, and
  recommended controls.
- Suggested frontmatter edits: description, invocation control, tools, model,
  context, paths, or arguments if relevant.
- Final decision: approve, approve with edits, request changes, reject, or
  escalate.

## Features

- Reviews `SKILL.md` frontmatter, body instructions, trigger text, and output
  expectations.
- Applies Claude Code skill behavior: descriptions load for discovery, full
  skill content loads on use, and skills can run inline or in isolated subagent
  context.
- Checks least-privilege tool scoping and user-only invocation for side-effecting
  workflows.
- Uses helpful-content questions to separate practical skill instructions from
  thin, duplicate, or unsupported copy.

## Use Cases

- Review a new project skill before merging it into `.claude/skills/`.
- Audit a plugin skill before distributing it to other repositories.
- Decide whether a skill should be automatically model-invoked or manually
  invoked by users.
- QA a skill library for overlap, vague descriptions, excessive context cost, or
  risky supporting files.

## Source Notes

- Claude Code skills are `SKILL.md` files with YAML frontmatter and markdown
  instructions. Descriptions help Claude decide when to load a skill, while the
  full body loads when the skill is used.
- Official Claude Code guidance distinguishes skills from subagents: skills are
  reusable instructions or workflows; subagents are isolated workers with their
  own context. A subagent can preload skills, and a skill can run in a forked
  context.
- The Claude Code skills docs document invocation control, tool restrictions,
  supporting files, dynamic context, subagent execution, and visibility
  overrides.
- Google's helpful-content guidance recommends self-assessment for reliability,
  usefulness, and "Who, How, and Why" clarity; this maps well to QA checks for
  public or team-distributed skill instructions.

## Duplicate Check

Checked `content/agents`, `content/skills`, the generated README catalog, and
open pull requests for `Agent Skill QA Reviewer Agent`, `skill QA reviewer`,
`SKILL.md review`, `Agent Skills QA`, `Claude Code skill review`, and the source
domains in the issue. Adjacent entries cover skill library curation and
enterprise skill governance, but no existing entry provides a concrete agent
prompt focused on QA review of an individual Agent Skill before adoption or
publication.

## Editorial Disclosure

Submitted as an independent source-backed HeyClaude content entry by `Desel72`.
It is based on public Claude Code and Google Search Central documentation. No
paid placement, referral, affiliate link, or vendor sponsorship is used.

## Sources

- Claude Code skills documentation: https://code.claude.com/docs/en/skills
- Claude Code features overview: https://code.claude.com/docs/en/features-overview
- Google Search Central helpful content guidance: https://developers.google.com/search/docs/fundamentals/creating-helpful-content