Will an agent actually load your skill?

An agent decides whether to use a skill from two lines of frontmatter. Skill Check reads your SKILL.md the way an installer does and lists what would stop it loading or triggering. Free, no sign-up, and your file never leaves your browser.

Check a skill.

or paste the file:

We ran it on 1,270 public skills.

Six well-known public skill collections, checked on 13 November 2026. A failure means the file breaks the Agent Skills format. A warning is a judgement call, and we say which below.

95had a hard failure: an invalid name, no description, or a description over the 1,024-character limit
1,013never say when to use the skill ("Use when…")
1,259never say what the skill is not for
1passed with no findings at all
How we counted, and what these numbers do not show

Each SKILL.md was run through the same code as the checker above, at these commits: anthropics/skills 2c5958e (20 skills), obra/superpowers 22f0b46 (15), vercel-labs/agent-skills 1a2c79d (9), wshobson/agents 9c61ab2 (185), K-Dense-AI/claude-scientific-skills 2d2065f (177) and ComposioHQ/awesome-claude-skills be2a406 (864).

  • 864 of the 1,270 come from one collection, so the totals lean on it. Of the 95 hard failures, 92 are there.
  • The two warnings are our house rules, not part of the format. A skill with no neighbours does not need a "Not for" line, and a skill can trigger well without the words "Use when".
  • The checker reads frontmatter with a strict parser. A value it rejects may still load in a forgiving agent.
  • These are good projects. Every skill we publish ourselves started with findings too.

What it checks.

The same checks as check_skill.py, the open validator inside tune-skill. A test compares this page's code with that script on 1,342 skill folders, and they agree on every one.

Fix it yourself, or hand it to us.

The checker finds what a script can find. The harder problems need a reader: a description that promises what the body does not do, two skills that answer the same request, a handoff with no format. That is what a tune-up covers.

Do it yourself

tune-skill

$0

The open skill we use for every tune-up. Your own agent runs it.

  • The full audit checklist, not only the scripted checks
  • Works in Claude Code, Cursor, Codex and any agent that reads a SKILL.md
  • You spend your own tokens and review the edits yourself
Install tune-skill
One skill

Tune-up

$19 per skill. Launch price, usually $29

We audit one skill by hand and send it back fixed.

  • Your SKILL.md returned as a ready-to-commit replacement file
  • Every finding with the quoted line, its severity and the smallest fix
  • Near-miss test requests for each boundary we changed
  • Small edits in your voice, not a rewrite
  • Back within two working days of payment
  • Full refund if we find nothing worth changing
Order a tune-up
A set that works together

Family tune-up

$99 up to 5 skills

For skills that share a repo or hand work to each other.

  • Everything in the tune-up, for each skill
  • Overlap check: each request gets exactly one owner
  • Handoffs between the skills given one shared format
  • A saved baseline so a later edit cannot quietly break a sibling
  • Back within three working days of payment
Order a family tune-up

Prices in USD, before any local taxes. Ordering opens an email to us; we reply with a secure payment link, and the clock starts when you pay. More than five skills, or private ones? See custom work.

Before you paste.

Where does my SKILL.md go?

Nowhere. The check is a script that runs in this page, and pasted text is never sent to us or stored. If you give a GitHub link, your browser fetches the public file directly from GitHub, or from jsDelivr's public mirror of GitHub when GitHub's lookup limit is used up. You can read the script at /check/check-core.js.

Does a clean result mean my skill will trigger?

No. It means nothing in the file's shape is stopping it. Whether an agent picks your skill for the right request depends on what the description says and on the other skills installed beside it, and no script can judge that. Test it on real requests.

I got warnings. Do I have to fix them all?

No. Fix every FAIL. A WARN is a question to answer: a skill with no neighbours does not need a "Not for" line, and a heading that differs from the name on purpose is fine.

What exactly do I get for $19?

One email with three things: your SKILL.md as a replacement file, a list of findings that each quote the line and give the smallest fix, and a short set of test requests that should and should not trigger the skill. We tell you which checks we ran and which we could not.

What is the refund policy?

If the audit finds nothing worth changing, we refund the full amount without being asked. If you are unhappy with what you get back, reply within 14 days and we refund you. Refunds go back to the original payment method.

Can you tune a private skill?

Yes. Paste the file into your order email with anything sensitive removed. We read it to do the work and do not publish or reuse it. We never need access to your repo, your systems or your API keys.

What about GitHub's rate limit?

GitHub allows 60 anonymous lookups an hour from one network address, which a shared office or mobile network can use up. When that happens the page finds the skill through jsDelivr's public mirror instead and the check carries on. One check is skipped on that route, whether the files the body points to exist, and the result says so.

Send us the skill that misfires.

Tell us what it should do and what it does instead. For one skill it is $19 at the launch price (usually $29), back in two working days, refunded if we find nothing to change.

[email protected]