Three checks before you install a Claude Code skill (from a review of 783 skills)

Before you install a Claude Code skill, open its SKILL.md and check three things.

1. Does it contain real steps? A one-line stub, a pointer to another file, or unreadable ciphertext means there is nothing to install.
2. Where does your data go? Search the file for URLs, keys and sign-up links. Calling an outside service is fine when you know about it and chose it.
3. Would it work on your machine? Hard-coded paths, someone else's Slack channel, instructions to switch off the sandbox, or steps for getting past a site's bot checks are all reasons to skip it.

These come from a review we ran on the AgentCraft directory in November. We collected every Claude Code skill on GitHub that we had yet to list, 783 in total, and dropped 317 of them. That is about two in five.

The first check alone would have caught 128 of the 317.

Stars told us almost nothing

Quick guess: did repo stars help predict which skills got dropped?

Barely. Dropped skills came from repos with a median of 1,572 stars. Kept skills came from repos with a median of 1,643. And 220 of the 317 dropped skills live in repos with more than 1,000 stars. Popular repos often ship demos, templates and test fixtures next to their real skills, and those were a large share of what got dropped.

How the 783 were collected and reviewed

1. Find repos. Repos linked from five large awesome lists of Claude skills, plus eleven GitHub searches on topics such as agent-skills and claude-skills: 1,888 repos.
2. Filter. At least 100 stars, no forks, no archived repos, pushed to in the last year, and no repos that only mirror other people's skills: 1,569 repos, of which 1,414 contain at least one SKILL.md.
3. Pick. A few skills per repo depending on stars (20 for 5,000+ stars, down to 2 for 100+), keeping the most starred copy when the same skill name shows up in several repos: 5,169 skills. 783 of them were new to the directory and came back from review in a usable form.
4. Review. Claude Haiku read each SKILL.md, eight per call, and either filed it under one of 38 categories or dropped it with a short reason.

The review prompt drops a skill when it exfiltrates data or secrets, harvests credentials, evades bot detection, anti-fraud or platform rules, is malware or offensive hacking against others, is mainly an ad or paid-API funnel, is a placeholder, template, test fixture or near-empty file, is a personal config useful only to its author, or is sexual content.

Why the 317 were dropped

A few examples:

• Demo. In a repo with about 68,000 stars, one skill only runs the date command and another fetches the temperature of one fixed city. Great for learning how skills work, useless to install.
• Paid API funnel. Seven near-identical skills each send you to sign up for the same paid transcript API and paste in a key.
• Unsafe. One skill sends the run's context to an outside service on its own. Nine of the 41 come from two security scanners, NVIDIA SkillSpector and Cisco's skill-scanner, which keep deliberately malicious skills as test cases, including a recipe assistant whose guardrail tells the agent to add cyanide. Those are test samples doing their job. Counting their harmless fixtures too, 25 of the 317 dropped skills come from these two scanners' test folders.
• Encrypted. Six skills from one repo are stored with git-crypt, so the SKILL.md you would install is ciphertext.

We kept 466. One was a duplicate, so 465 went into the directory.

Limits worth knowing

Each skill got one pass from one small model, reading only the first 1,800 characters of its SKILL.md.

We also got an accidental stability test. A name-matching bug sent 266 skills that were already listed through review a second time. Same prompt, same model: 257 kept, 9 dropped. Those 266 are left out of every number above.

A dropped skill also says nothing bad about its repo. Anthropic's own skills repo has a template skill that was dropped here simply because it gives an agent nothing to do.

The full write-up, the exact review prompt and a JSON file with all 317 reasons (names removed) are on the original post:

https://agentcraft.dev/en/blog/claude-code-skills-review-2026-11/

The 465 kept skills are in the AgentCraft skills directory (5,101 skills, with install commands):

https://agentcraft.dev/en/skills/

Disclosure: I'm Adrian Vale and I help run AgentCraft.