vetting-a-skill-before-install
Skill1 starsChanged 3 months agoApache-2.0View on GitHub
Use before installing, importing or enabling a third-party skill, plugin, marketplace or skills repository in Claude Code, Codex, Antigravity, Cursor or any agent. Pins the source to a commit, audits every skill and the agent configuration it ships, reads what the audit could not cover, and ends in install, mirror or refuse. Not for auditing this registry's own skills.
- Pipes a download into a shell
In this repository19 other files it gives its agents.
- AGENTS.mdAGENTS.md
- agent-loop-designSkill
- ai-supply-chain-securitySkill
- anti-slop-pr-and-writingSkill
- authoring-portable-skillsSkill
- brainstormingSkill
- cross-language-portSkill
- giving-code-reviewSkill
- handoffSkill
- incident-responseSkill
- pr-review-and-releaseSkill
- receiving-code-reviewSkill
- refactoring-safelySkill
- systematic-debuggingSkill
- test-driven-developmentSkill
- using-agtmlsSkill
- verification-before-completionSkill
- web-research-and-source-triageSkill
- writing-plansSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 28
Keep tool calls few: every call re-sends the conversation so far.
- Very specific: line 44
Read what runs unasked, in one call: Read every file the audit lists as NOT AUDITED at once: one cat of them all, or all the reads in a …
- Very specific: line 52
Install: no CRITICAL or HIGH finding, every NOT AUDITED file read, nothing runs unasked that you would not run yourself.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Vetting a skill before install
Trigger. Someone wants to add a skill, plugin, marketplace or skills
repository this registry did not produce: /plugin install, a clone into
.claude/skills or .agents/skills, import-skill, or "can I trust this
repo?". A skill runs with your permissions: its text is read as
instructions, its hooks run on events you did not trigger, its scripts run
in your shell. Install counts are not a control.
Keep tool calls few: every call re-sends the conversation so far. Start with the audit, which answers the first questions for you.
Steps
- Audit.
agtmls audit --foreign <path, or url@commit>, with--format jsonto parse it. A CRITICAL or HIGH finding fails. Its coverage section states the source, and lists everyNOT AUDITEDfile: agent configuration it did not judge (hooks.json,settings.json,.mcp.json,plugin.json,CLAUDE.md,AGENTS.md) and scripts outside any skill. Silence about a file is not a pass. - Pin. Take the 40-hex commit from the audit's
source:line; a branch or tag can move between the audit and the install. If it says the source is not a git checkout, it cannot be pinned: count that against it rather than searching further. Check the owner and name against the project you meant; note the licence. - Read what runs unasked, in one call. Read every file the audit
lists as
NOT AUDITEDat once: onecatof them all, or all the reads in a single turn. Look for hooks that fetch and run code or send data out (curl | sh, a transcript upload), MCP servers not pinned to a version or able to run any command, and checksums fetched from the same place as the download. Say what reading cannot show, such as what a download does when it runs. - Decide.
- Install: no CRITICAL or HIGH finding, every
NOT AUDITEDfile read, nothing runs unasked that you would not run yourself. Install the audited commit, not the branch. - Mirror: worth having, but you want control of updates. Copy it at the audited commit; re-audit before taking an upstream change.
- Refuse: a finding you cannot explain away, hooks or scripts you have not read, or a source you cannot pin.
- Install: no CRITICAL or HIGH finding, every
An install through agtmls is recorded, and agtmls verify shows later
drift. A plugin manager's install is not: audit again before each update.
Report
Open with the source: the commit you would install, or that it cannot be
pinned and so cannot be installed as audited. Then the audit command and its
counts, what you read by hand, what could not be checked, and the decision
with its reason. reference.md has worked outcomes and the red flags.
When not to use
- Auditing this registry's own skills:
agtmls audit --all. - Reviewing ordinary code changes: that is code review.
- The source is already mirrored and pinned, and unchanged since its audit.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

