3,515 skills scanned · 80% can't be dispatched

Your skill works. The agent just doesn't know when to use it.

We scanned 3,515 public skills. 91% had zero critical issues, and 80% shipped a description an agent can't act on: no trigger conditions, no statement of when to use it. Yours is probably one of them. SkillCheck tells you, locally, in seconds. Free to start; Pro catches what ships bugs.

Check my skill for free See what Pro catches
Built for Claude Code· Open standard· Runs 100% local· v3.30.0
Score distribution · public skills
6.2%
19%
61%
14%
0%
0–39
Poor
40–59
Weak
60–79
OK
80–99
Good
100
Excellent
80%
missing proper descriptions
0%
scored Excellent

You built a skill. But is it any good?

Skills are an open standard now. Yours runs in Claude Code, Cursor, VS Code and a dozen agents, and every one of them is a chance to look broken.

You won't know until someone files an issue. Or you run SkillCheck.

01 Does the YAML frontmatter follow the spec?
02 Will it trigger when users actually need it?
03 Is it accessible, readable, and free of LLM slop?
04 Could it over-grant tools or leak a credential?
Tested at scale

We scan the public skill ecosystem on every release

As SkillCheck adds checks, the bar gets stricter. The scores reflect a tougher standard, not worse skills. What the corpus shows is not broken files, it is undiscoverable ones: 4.4 points separate vendor repos from everyone else, and three quarters of the skills published by Anthropic, Microsoft, Google, GitHub and NVIDIA carry the same unusable description. Hasan et al. found the same failure in 97% of 856 MCP tools.

3,515
skills scanned
64
average score / 100
80%
missing descriptions
91%
zero critical issues
avg score over time: v3.12 avg 65 → v3.18 avg 58 → v3.20 avg 57 → v3.21 avg 57 → v3.26 avg 63 → v3.28 avg 64 repos scanned: 8 → 30 · Pro tier · last scan 03-08-2026
What it checks · 28 categories

Free catches the obvious. Pro catches what ships bugs.

Every check runs locally. No skill content ever leaves your machine. Start free; the upgrade is the difference between “valid” and “safe to ship to other people”.

Free · 6 areas $0

Schema, structure and surface quality. Open source (MIT), runs as a skill file, no install.

Structure
Missing fields, invalid names, broken YAML, XML injection, effort / maxTurns / disallowedTools
Body & Naming
Content requirements, formatting, anti-pattern lint, conventions, reserved words, gerunds
Semantics
Contradictions, ambiguity, platitudes, workflow-in-description, railroading
Quality Patterns
Examples, error handling, triggers, output format, prerequisites, gotchas
OWASP Agentic (Cat 26)
8 deterministic checks: tool over-grant, identity override, unpinned deps, eval/exec & more
Discovery & Density
Finds skill candidates in your configs; flags hollow gotcha sections
Pro · adds 10+ $79

The judgment checks: security, slop, accessibility, agent-readiness and governance across skills, plugins and MCP servers.

Anti-Slop
“Let's dive in”, hedge words, verb inflation, promotional adjectives, vague attributions
Security
PII detection, credential safety, path traversal, Bash guardrail validation
Visual & WCAG
Colour contrast and accessibility for skills that produce UI
Token Budget
Context efficiency, budget analysis, frontmatter overhead
Enterprise
Hardcoded paths, env config, audit support, data-storage path safety
Agent Readiness
28 checks across 6 pillars · L0→L3 maturity scoring
Workflow
Exit criteria, unbounded iteration, deploy-without-verify (Trail of Bits catalog)
Trigger Collision & Eval Kit
Cross-skill overlap detection; auto-generated should / should-NOT test prompts
Plugin & MCP governance
Cat 24 manifest governance + Cat 23 identity-override, annotations, CLI literacy
OWASP graders (Cat 26)
Deeper intent checks your own AI judges: goal hijacking, memory poisoning. No API key
Repo Maturity (Cat 27)
Reads git history: commit count, recency, single-sitting; flags skills written once and never used
Category Tier What we catch
How it works

Two ways to run it

Free

Claude reads a skill file and runs checks inline. No install, no binary.

1 Copy skill-check into ~/.claude/skills/
2 Say “skillcheck my skill” in Claude Code
3 Get instant inline feedback
Pro

A local Go binary that connects to Claude Code as an MCP tool server.

1 Download the binary from your purchase email
2 Add to ~/.claude.json with your license key
3 All Pro checks unlock automatically
CLI · validates three artifact types
SKILL.md
skillcheck check \
 ~/.claude/skills/my-skill
Plugin manifest
skillcheck check-plugin \
 --pro .claude-plugin/plugin.json
MCP server
skillcheck check-mcp \
 --pro tools-list.json

--pro flags need SKILLCHECK_LICENSE. Or paste any of the three into the browser validator, runs entirely client-side via WebAssembly.

Pricing

One purchase. Every Pro check. Forever.

No subscription, no seats, no API keys. $79 once unlocks all 28 categories on macOS, Linux and Windows. Want to try Pro on a single project first? A 30-day Pro Pass is $15 at the same checkout. Prices include tax, so what you see is what you pay.

Free

$0 forever
  • Structure validation
  • Body & naming checks
  • Semantic consistency
  • Quality patterns
  • OWASP deterministic (Cat 26)
  • No install needed
  • Open source (MIT)
Get started
Most popular

Pro

$79 lifetime · tax included
  • Everything in Free
  • Anti-slop detection
  • Security scanning
  • Token budget analysis
  • WCAG accessibility
  • Enterprise readiness
  • Agent Readiness (L0–L3)
  • Plugin manifests (Cat 24)
  • MCP tool lists (Cat 23)
  • OWASP graders (Cat 26)
  • Repo Maturity (Cat 27)
  • Go binary & MCP server
Get Pro or try a 30-day Pro Pass · $15
Coming soon

Team

$49 /mo · up to 10
  • Everything in Pro
  • Custom rulesets
  • Team dashboard
  • CI/CD integration

Pro includes a GitHub Actions integration: uses: olgasafonova/skillcheck@v3 blocks low-quality skills in CI.

Validated skills

Every passing skill gets a public report

A score, a grade and badges: proof your skill is safe to depend on. Here are some that scored a perfect 100.

Browse all reports
anthropics/ Excellent
web-artifacts-builder
100 / 100
microsoft/ Excellent
github-issue-creator
100 / 100
microsoft/ Excellent
podcast-generation
100 / 100
Shipping constantly

Recent releases

Full changelog
v3.30 Aug 2026 Check skills from anywhere: a new skillcheck serve command answers web requests with findings and a score, whole GitHub repos get checked in one go, the MCP server works over the network, and a hosted server can require each caller's own Pro license.
v3.29 Aug 2026 A sturdier MCP server: new official foundation from Anthropic, and a crashing check can no longer take the whole server down. Everything you have set up keeps working.
v3.28.1 Jul 2026 Corrects the license activation instructions bundled in the Pro download, which showed a key format Lemon Squeezy does not issue. No change to the program itself.
v3.28 Jul 2026 Four more anti-slop patterns (Cat 5) from an audit against Wikipedia's "Signs of AI writing": knowledge-cutoff disclaimers, chat-UI citation artifacts left in pasted text, assistant correspondence, and copula inflation.
v3.27 Jul 2026 Anti-slop cluster-mining patterns (Cat 5): AI vocabulary with co-occurrence escalation, significance inflation, cross-sentence antithesis, and corporate/internet cliché lists, mined from four community anti-slop repos.
v3.26 Jul 2026 Reference-aware checks expand to composability (Cat 17), observability (Cat 18), and knowledge density (Cat 22): input/output contracts, confidence signals, and knowledge checks now follow a skill's references/ files too.
v3.25 Jun 2026 Reference-aware checks: content checks now read a skill's references/ files, so a lean SKILL.md with detail moved into references/ is no longer flagged for "missing examples."
v3.24 Jun 2026 Restraint (Cat 28): warns when a skill says "keep it simple" but never says to keep validation, security, and accessibility; Pro rewards "simple but safe."
v3.23 Jun 2026 Fewer false positives: the frontmatter XML check and the reviewer-pattern classifier no longer fire on placeholder tokens and action skills.
v3.22 Jun 2026 Repo Maturity (Cat 27): reads a skill's git history to flag skills written once and never used; Pro, scored in its own lane.
FAQ

Questions, answered

No. SkillCheck is an independent project. Free validates against the Agent Skills open standard created by Anthropic; Pro adds extra quality checks like anti-slop, WCAG and security scanning.

Built and tested for Claude Code. Because it follows the open standard and MCP protocol, it works with Cursor, VS Code, Windsurf and other compatible tools.

No. Every check runs locally, in the CLI, in CI, or in the browser WASM build. Nothing is uploaded, which matters when your skills hold confidential product knowledge.

A lifetime Pro license: the Go binary, MCP server, and all Pro checks across skills, plugins and MCP servers, including future categories. No subscription, no seats. If you want to try Pro on one project first, the 30-day Pro Pass is $15 at the same checkout: same binary, same checks, and the license key expires 30 days after purchase.

The binary drops back to the Free checks automatically. Nothing breaks, nothing renews, and nothing is charged. Buy a new pass or upgrade to lifetime whenever you need the Pro checks again.

Yes. Pro ships a GitHub Actions integration. Add uses: olgasafonova/skillcheck@v3 to block PRs that introduce low-quality skills.

SkillCheck reports issues; you decide what to fix. Some checks are suggestions, not requirements.

Ship with confidence

Find the problems before someone files an issue.

Check my skill, free Get Pro · $79 lifetime

30-second free check · no signup · runs in your browser