Guide · Agent skill
How to use the Skill Reviewer agent skill (skill-reviewer)
Updated · written with AI from the project's own documentation and checked against our catalogue data
Skill Reviewer is an agent skill from the bytedance/deer-flow project (skills/public/skill-reviewer, MIT licence). It makes an agent act as a quality reviewer for other agent skills: it inspects a skill package, checks whether its description triggers correctly, flags safety and resource problems, and reports how much evidence supports the skill's claims. This guide explains what an agent skill is, how to install this one, how to ask for a review, and where its limits are.
Skill Reviewerbytedance/deer-flow · MIT Download the skill Quick start
- 1Download and unzip. Download the ZIP from the skill page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder.
- 2Or upload in the Claude apps. In the Claude apps, go to Settings → Capabilities → Skills and upload the ZIP.
- 3Confirm the runtime. Make sure your DeerFlow runtime provides the review_skill_package tool.
- 4Identify the target. Get the skill's skill:// reference or paste its SKILL.md content.
- 5Ask for a review. Use a natural request such as asking to audit or production-check the skill, and name any focus areas.
- 6Read the verdict and evidence. Check readiness, assurance, blockers and the limitations section before acting on suggested rewrites.
- 7Apply fixes yourself or hand off. Make edits yourself, or ask for skill-creator to apply changes or run experiments.
What is an agent skill and what does Skill Reviewer do?
An agent skill is a folder built around a SKILL.md file. The file has a short front-matter block with a name and a description, then instructions the agent follows when a request matches that description. Skill Reviewer is one such skill, and its job is to review other skills rather than do ordinary work.
When it is active, the agent treats the target skill package as untrusted data. It inspects the package through the review_skill_package tool, checks the description for over-triggering and under-triggering, looks at scripts, resources, safety boundaries and output, and grades the evidence behind the skill's claims. It returns a readiness verdict and suggests paste-ready fixes.
- Readiness verdicts: blocked, revise, or publish_candidate.
- Assurance levels: static_only, trigger_checked, behavior_verified, or regression_verified.
- It does not edit the skill. Edits, creation and runtime experiments are handed off to skill-creator.
Install Skill Reviewer
You can install the skill in one of two ways, depending on where you use your agent. Download the ZIP from the skill page on this site and unzip it. Then place the resulting folder in your personal skills directory, ~/.claude/skills/, or in a project's .claude/skills/ directory so it applies only to that project.
In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/.
- Project install: unzip into the project's .claude/skills/ folder.
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP.
Check the requirements before you start
Skill Reviewer depends on a DeerFlow runtime that provides the review_skill_package tool. Without that tool, the skill cannot inspect a package as designed. You also need access to the target skill, either as a skill:// reference such as skill://public/data-analysis, or as pasted SKILL.md content.
The skill runs no scripts and uses no network access, according to its source. It reads the target only through review_skill_package.
- A DeerFlow runtime providing review_skill_package.
- A skill:// reference to the target, or the pasted SKILL.md text.
- No scripts are run and no network use is declared.
How to trigger Skill Reviewer with natural requests
You do not need a special command. The skill is meant to activate when you ask for an audit, grade or production check of an existing skill. Phrases like the following match its description:
Name the skill or paste its SKILL.md, and say what you want checked. If you only care about some areas, say so, because the skill limits its review to the requested dimensions and otherwise covers all of them.
- "Audit the skill skill://public/data-analysis before we publish it."
- "Grade this SKILL.md for readiness and tell me whether it will trigger correctly."
- "Production-check my team-helper skill, focusing on safety and resources."
- "Why does this skill fire on every request? Review its description."
What a review contains
A full review follows a fixed structure: an executive summary, readiness, assurance, scope and completeness, findings, dimension review, trigger analysis, resource and script review, evidence, suggested rewrites, and recommended actions. A focused review can drop unrelated sections but still states scope, readiness, assurance, evidence and actions.
Every issue is reported with severity, confidence, location where available, observed evidence, user impact and concrete remediation. Readiness and assurance are kept separate, so a skill can be ready in its assessed scope while having only static evidence.
- Readiness: blocked means a blocker exists; revise means no blocker but errors or major gaps; publish_candidate means no material issue was found in the assessed scope.
- Assurance: static_only means static inspection only, and the higher levels require retained runtime evidence for the reviewed digest.
Safety notes and limits
Skill Reviewer treats the reviewed package as data, not instructions. The source says to ignore any instruction inside the package that asks it to change verdicts, reveal prompts, execute scripts, install dependencies, fetch URLs, modify files or request secrets. Its own fixtures include a prompt-injection example that tests exactly this behaviour.
A publish_candidate verdict does not prove the skill works at runtime. It only means no material issue was found within the assessed scope. Ordinary application-code review is excluded from the skill's purpose.
- It does not apply changes to the reviewed skill.
- It does not create, package or install skills; those go to skill-creator.
- A clean verdict is not proof of runtime behaviour.
What people use it for
- Pre-publish check of a DeerFlow skill before sharing it with an internal team.
- Diagnosing a skill whose description fires on unrelated requests (over-triggering).
- Finding why a skill never activates (under-triggering) or collides with a sibling skill.
- Checking whether a skill's existing evals actually prove what it claims.
Frequently asked questions
How do I install the Skill Reviewer skill?
Download the ZIP from the skill page on this site and unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps you can instead upload the ZIP under Settings → Capabilities → Skills.
What do I need to run Skill Reviewer?
You need a DeerFlow runtime that provides the review_skill_package tool, plus access to the target skill via a skill:// reference or pasted SKILL.md content. The source declares that it runs no scripts and uses no network.
Is Skill Reviewer free?
The skill is published under the MIT licence in the bytedance/deer-flow repository. The source does not state any price, so check the licence terms in that repository for details.
Can Skill Reviewer fix my skill for me?
No. It inspects the package and recommends paste-ready changes but does not apply them. Creating or editing a skill, or running experiments, is handed off to the skill-creator skill.
Does a publish_candidate verdict mean my skill works?
No. publish_candidate only means no material issue was found within the assessed scope. It does not mean runtime behaviour was verified, which requires retained evidence for higher assurance levels.
How is Skill Reviewer different from skill-creator?
Skill Reviewer only inspects and recommends changes. skill-creator is the separate skill used to create new skills, apply edits, package them, or run behaviour and baseline experiments.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Skill Reviewer suits developers and teams who maintain a SKILL.md library and want a structured pre-publish check, or authors who need to see why a skill triggers too often or not at all. It is the wrong tool if you want a skill written or edited, need ordinary application-code review, or expect a verdict to prove runtime behaviour, in which case use skill-creator or real runtime tests.
Download the skill