- Home
- Skills
- Prompt & Skill Engineering
- skilllint - Vet Agent Skills Before You Install
More screenshots
Works with the AI Agents you already use
skilllint - Vet Agent Skills Before You Install
Private lint of agent skills: structure, triggers, secrets, injection, broken refs. Offline, verdict + SHA-256.
$19
skilllint - Vet Agent Skills Before You Install
Example session with this skill installed
Lint the SKILL.md at ./agents/weather-fetcher/SKILL.md and tell me if it is safe to install.
- Read your context and instructions
- Compiled the skilllint - vet
Summary
The candidate skill xmtplabs/xmtp-agent-examples was linted and failed. It has no
frontmatter, so there is no description for an agent to match against and no reliable way
to know the skill exists or when to reach for it. The package may still be good
documentation - as an agent skill it is unreachable.
Verdict
| Metric | Value |
|---|---|
| Verdict | FAIL |
| Score | 50/100 |
| SHA-256 | b9757a1d905df1061b2662fe2cc148c8335ee8d43f876c3a065b0db9956f55be |
| File size | 1,633 bytes |
Findings
R-001(critical) - frontmatter (--- ... ---) not foundR-002(high) - name is missing or malformed: ''R-003(high) - description is missingR-008(medium) - referenced file not found: skills/building-agents/
Raw output
{
"skill": "xmtplabs__xmtp-agent-examples__skills_SKILL",
"path": "/tmp/opencode/skills-check/xmtplabs__xmtp-agent-examples__skills_SKILL/SKILL.md",
"score": 50,
"verdict": "FAIL",
"sha256": "b9757a1d905df1061b2662fe2cc148c8335ee8d43f876c3a065b0db9956f55be",
"bytes": 1633,
"findings": [
{
"rule": "R-001",
"severity": "critical",
"message": "frontmatter (--- ... ---) not found",
"line": null
},
{
"rule": "R-002",
"severity": "high",
"message": "name is missing or malformed: ''",
"line": null
},
{
"rule": "R-003",
"severity": "high",
"message": "description is missing",
"line": null
},
{
"rule": "R-008",
"severity": "medium",
"message": "referenced file not found: skills/building-agents/",
"line": null
}
]
}
What the agent does with this
The verdict is FAIL, so the skill is not installed. The four rule ids are reported back to
whoever sent the candidate, because a rejection with rule ids is actionable for the author
and a rejection with adjectives is not. The SHA-256 is kept: re-lint the same bytes later
and the verdict is identical, and a changed hash means the skill is not the one that was
vetted.
The check was local: the file was read on this machine, nothing was uploaded, and the
verdict carries the hash of exactly the bytes that were checked.
Connects securely to your tools. The creator never sees your data.
About this skill
A skill is an instruction set that gets injected into an agent's context, and three failures dominate - none of them announces itself.
It never fires: no trigger wording in the description, so the agent never reaches for the capability. Nothing errors, you just never get it.
It fires everywhere: a vague description plus a 100 KB file means every relevant request pays for the whole thing.
It carries something: a secret literal, an injection marker, or "act without asking".
This skill makes that check local, private and reproducible. The candidate's bytes are read on your machine, nothing is uploaded anywhere, and the verdict comes with the SHA-256 of exactly what was checked.
What the check covers:
- structure: frontmatter, name, description, trigger wording
- consistency: dead instructions (TODO/deprecated/placeholder), references to files that are not in the package, files over 60 KB
- safety: literals that look like secrets, prompt-injection and exfiltration markers, instructions that grant unbounded autonomy
The verdict is PASS / WARN / FAIL with a score, the findings (rule id, severity, message) and the sha256 of the bytes that were checked - so the same file always produces the same verdict, and a changed file changes the hash. The hash is the input for a signed receipt and for an allowlist.
When several candidates are in play, compare_skills ranks them by verdict, then score, then size, and shows the evidence side by side.
Runs fully offline with the bundled script (standard library only). The hosted MCP endpoint is optional, not required.
-
SKILL.md - the check loop the agent follows
-
reference/rules.md - the 11 rules, what fires them, how to fix each
-
reference/setup.md - local and hosted connection, how to read a verdict
-
reference/candidates.md - comparing candidates, what to ask the author
-
scripts/linter.py - the linter itself, standard library only
-
Static analysis: it reads bytes, never executes the skill and never runs its scripts. A skill can pass the lint and still be wrong.
-
PASS means "structure is sound", never "safe". Eleven rules are an opinionated minimum, not a threat model.
-
The "referenced file not found" check only means something when the whole package is present: lint the directory, not a lone file.
-
The linter is opinionated and argues with itself in public: we lint our own skill with it, and we publish both our findings and our own false positives.
Static analysis only: it reads bytes, never executes the skill and never runs its scripts - a skill can pass the lint and still be wrong.
PASS means 'structure is sound', never 'safe': eleven rules are an opinionated minimum, not a threat model.
The 'referenced file not found' check only means something when the whole package is present - lint the directory, not a lone file.
The linter argues with itself in public: we lint our own skill with it and publish our own false positives.
The checker decides; it never installs anything and never modifies the candidate.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 27 days ago
Requires
Permissions
File Scopes
The skill reads a candidate SKILL.md from disk and writes a verdict to stdout. It does not write files, does not open a browser and makes no network calls. The hosted MCP endpoint is optional - everything works offline.