Agent Skills

Checks and tools

Every skill listed on qai.io/skills is read at an exact commit and checked three ways. These are the checks, and the tools that run them on your own skill.

Validate your SKILL.mdPaste it or drop the folder, and see what each agent will reject, with the fix. It runs in your browser.Open Validate

The format check

The format check reads SKILL.md the way each agent does, as Agent Skills spec, Claude Code, Claude apps & API upload, OpenAI (Codex & API) and VS Code & Copilot, and says for each whether the skill passes, passes with warnings or fails, with the rule and its fix. How agents differ explains each.


The pattern check

The pattern check reads every text file in a skill’s folder, at the commit its page names, against 45 published rules for things worth a second look, such as download-and-run commands, hidden text and credential access. It doesn’t run the skill or read the web pages it points to, and it can’t judge everything its instructions might lead an agent to do. Each skill gets one of three results:

ResultWhat it means
Nothing flaggedThe check found nothing serious in the skill’s files.
Items to reviewThe check found something worth reading before you install, or couldn’t read part of the skill. Its page shows the lines it found.
Held for reviewThe check found something serious. The skill isn’t listed and has no page.

A clear result isn’t proof that a skill is safe, harmless, unchanging or any good. A finding doesn’t mean bad intent either: many good skills run commands, reach the network or read keys they need. How Skills works lists every rule.


What it can touch

This is how much a skill can do, not whether it is risky. A skill with none of these shows Instructions only, though its text can still ask the agent to run commands.

LabelWhat it means
Runs shell commandsIts instructions or files give the agent commands to run on your computer.
Reaches the networkIt sends or fetches data over the internet, for example with curl, an HTTP library or a package install.
Names web addressesIt mentions outside web addresses the agent may visit. We list them but don't follow them.
Bundles scriptsThe folder holds code files the agent may run.
Pre-approves toolsIts allowed-tools field lists tools the agent may use without asking you first, in agents that honour it.
Runs commands as it loadsClaude Code runs its `!command` lines when the skill loads, before the agent reads it, if your permission rules allow them.
Adds hooksIts frontmatter registers commands that run automatically on agent events for the rest of the session.
Touches credentialsIt reads, asks for or contains keys, tokens, passwords or similar secrets.
Downloads softwareIt fetches programs or packages from the internet, or from a package source other than the default.
Follows outside instructionsIt tells the agent to fetch text from the web and act on it. That text can change after our check.
Sets up background tasksIt schedules jobs or startup items that can keep running after you remove the skill.
Installs more skillsIt tells the agent to install other skills, plugins or tool servers.
Changes agent settingsIt writes to agent memory or settings files such as CLAUDE.md, AGENTS.md or .claude/settings.json.
Turns off system protectionsIt removes download warnings, adds antivirus exclusions, switches off a firewall or loosens permissions.
Deletes dataIt includes commands that delete whole folders, wipe disks or overwrite devices.
Handles moneyIt handles wallets, seed phrases, token approvals, trading keys or referral links.
Includes files we can't readIt ships binaries, archives or documents whose contents this check cannot look inside.

A skill’s page shows what it can touch at a glance to everyone. With a free account it also shows the file and line behind each one.


The QAI Index

A score from 0 to 100 for comparing skills, from public signals and worked out again at every refresh. It is a guide, not a verdict: a higher number doesn’t mean a skill is checked more closely, safer or better for your job.

PartPointsWhat it counts
AdoptionUp to 40GitHub stars on the repo that holds the skill, shared among its skills, and copies of the skill in other repos
QualityUp to 30Our pattern check, the format check, a named licence and a description of a sentence or more
UpkeepUp to 20How recently the skill's folder changed
TrustUp to 10An official collection, or an organisation that has confirmed its domain with GitHub

Tools for your own skill

  • Validate a SKILL.md: paste a file or drop a folder to see what each agent will reject, with the fix, or build one from a form. It runs in your browser and uploads nothing. Open to everyone.
  • Check a repo: paste a public GitHub repo to see what each SKILL.md in it can touch, what the pattern check flags and which agents would reject it, all in your browser. Needs a free account.
  • Get listed: send a public repo with a SKILL.md that has a name and description. The form runs both checks first, then we review the repo for a coming refresh. Needs a free account.
  • QAI Index badge: on any skill’s, repo’s or publisher’s page, a badge for a README or a website that shows its QAI Index and links back to the page. Open to everyone.
  • Watch: on a skill’s page, get an e-mail when its check changes. Needs a free account.
  • Stacks: keep the skills you use together and install them with one command per agent. Needs a free account.

A free account needs only Google or an e-mail link, with no password.