A Claude Code skill is mostly a Markdown file — but once installed, it runs with your tools, reads your code, and can tell the agent to execute things. I install skills from community collections all the time, and before any of them lands in ~/.claude/skills, I run the same 30-second audit. Here it is, with three real skills as worked examples.
All three examples below come from laolaoshiren/claude-code-skills-zh (445+ curated skills, 21 original ones; companion site: https://claude-skills.bt199.com). I picked them because each one passes the audit in a different instructive way.
Check 1: Read the frontmatter — when does it fire?
Every skill starts with a YAML frontmatter block: name and description. The description is not decoration — it is the trigger. Claude Code reads it to decide when the skill activates, so a vague description means the skill fires at the wrong time (or never).
Good:
--- name: zh-code-reviewer description: "中文代码审查专家 - 生成详细的中文代码审查报告" --- Paired with an explicit trigger section in the body ("activates when the user asks for a code review, code check, or quality assessment"), this tells me exactly when it will wake up. If a skill's description is a marketing slogan instead of a trigger ("the ultimate best reviewer!!!"), I skip it.
Check 2: List the bundled files — is it just Markdown?
A skill directory can contain more than SKILL.md: scripts, hooks, and extra files that run code. That is not automatically bad, but it changes the trust question from "what does this text say?" to "what does this code do?"
All three of my examples ship as a single SKILL.md — no scripts, no hooks, nothing executable bundled:
-
skills/zh-code-reviewer/—SKILL.mdonly -
skills/zh-readme/—SKILL.mdonly -
skills/api-tester/—SKILL.mdonly
If a skill does bundle scripts, read them. A 20-line shell script is auditable in a minute; a minified blob is a reason to walk away.
Check 3: What tools does it reach for?
A skill can only act through the tools its instructions invoke. Read the workflow section and note the verbs.
zh-code-reviewer is explicit: step 1 is "read the target files using Glob and Read". Read-only tools, a clearly bounded job, and the output is a structured Chinese report. Nothing in it sends requests or writes files.
api-tester is the more interesting case, because API testing can touch the network. Its description draws the authorization line up front: it distinguishes "generating tests" from "executing requests", does not guess status codes or response shapes, and does not send write requests in an unconfirmed environment. That is the pattern to look for — a skill that names its own boundary is one you can trust to respect it. A skill that never mentions authorization at all is the one to be careful with.
Check 4: Does it promise real evidence, not guesses?
The last check is about output quality, and zh-readme states it as a principle: "read the code and directory first, then write the README; don't guess the functionality from the project name." Its rules demand real commands, real paths, and real examples, and require keeping whatever is already correct in an existing README instead of overwriting it.
That principle generalizes: a skill that tells the agent to verify before writing is worth installing. A skill that tells the agent to invent plausible-looking content is not — no matter how polished its template looks.
Installing a skill that passes
If a skill passes all four checks, installation is the easy part. With Claude Code 2.1+, the collection above installs as a plugin:
/plugin marketplace add laolaoshiren/claude-code-skills-zh /plugin install zh-skills@claude-code-skills-zh Or the manual way:
git clone https://github.com/laolaoshiren/claude-code-skills-zh.git && cp -r claude-code-skills-zh/skills/* ~/.claude/skills/ The 30-second version
- Frontmatter: does the description state a concrete trigger?
- Bundled files:
SKILL.mdonly, or scripts you have actually read? - Tools: read-only verbs, or network/write actions with a stated authorization boundary?
- Output contract: verify first, or guess and polish?
Skills are powerful precisely because they run with your permissions. Thirty seconds of reading before installing is the cheapest security review you will ever do.
