Which skills we cover

We start from the most installed and most starred skills (data from Best Skills by LinklyAI, CC BY 4.0, which aggregates skills.sh, ClawHub and GitHub), add skills people search for by name, and add deep domain skills we found ourselves. We skip archived repositories and repositories whose install counts look inflated (see security).

Depth score (1-5)

Depth measures how much expertise ships with the skill, judged from the repository file tree and the files themselves:

1. a short prompt;
2. solid instructions, nothing else;
3. instructions plus reference documents, templates or helper scripts;
4. substantial references and scripts: a real procedure;
5. expert-grade: references, scripts and tests or evals, something you can leave running unattended.

A small skill can be excellent at its job. Depth tells you what you are getting, not whether it is good.

Verdicts

• Recommended: does what it says, sensible depth for its purpose, maintained, no real security concern.
• With caveats: useful, but you need to know something first: a paid dependency, heavy setup, telemetry, an unclear license or overlap with other tools.
• Skip: abandoned, misleading, a real security concern, or a thin prompt that adds nothing over asking the model directly.

Security status

The security status is separate from the verdict: a skill can be useful and still need broad access. For each skill we record facts with the file they come from: where it sends data, which credentials or browser data it touches, what it pre-approves, what it installs, whether it loads instructions from a server, and what we could not review. The status is then computed by a fixed rule:

• No issues found: we had every file in scope (the whole repository, or the skill folder for single-skill pages), read every script, and recorded nothing beyond running its own local scripts.
• Review before installing: anything worth knowing (network calls, an API key for the skill's own service, telemetry with an opt-out, a curl | sh installer, code outside the repository) or a partial review of a large pack. Each page lists the reasons and what we did not cover.
• Broad access: at least one of: sends your keys or tokens to another service, uploads your data by default, loads instructions from a server at run time, uses your browser session, pre-approves broad shell commands, installs hooks without asking, pushes or merges without asking, turns off safety prompts, telemetry without an opt-out, or obfuscated content.

Every skill is also scanned with NVIDIA SkillSpector in static mode, and we use its hits as leads. Raw pattern scans are very noisy: in our last run the scanner's own recommendation was "do not install" for 199 of 293 popular skills (68%), mostly because instructions mention commands like git push or rm -rf. Skills that look malicious are left out of the catalog entirely.

Who writes the reviews

Reviews and security facts are produced by AI reviewers working from each repository's files, file tree and scan results under a fixed rubric; every claim is limited to what those files show, and statuses are computed by code rather than chosen by the reviewer. If a review is wrong, tell us and we will fix it.

Updates

Each review and security check shows the date it was made. Repository numbers (stars, installs, last push) come from GitHub and skills.sh at the time of the last data refresh.