How we
review.

Fieldbook lists fewer skills on purpose. A skill appears here only after it passes the checks below, and it earns the Read in full mark once a person has read every instruction file.

What we look for

The checks we run

Every listed SKILL.md is fetched and checked each day for valid frontmatter and for patterns drawn from published research on malicious skills:

A match is not a verdict. A person reads it: a defensive mention, such as a skill that lists the commands it blocks, is cleared with the reason shown on the page. A new, unread match pauses the listing until it is read.

What the labels mean

Labels come from reading the skill, not from running it. A skill with no labels is guidance only.

Reviewed, and read in full

Every listing is reviewed: licence confirmed, SKILL.md scanned, access labelled, and a Fieldbook note written from its instructions. Read in full is a stronger mark: a person has read the skill and every file it tells the agent to load, end to end, and the date is shown. Submissions are read in full before they are listed.

Where the numbers come from

Install counts are the public counts from skills.sh (opens in a new tab), collected from the open-source Skills CLI. Fieldbook takes a snapshot every day; "today" is the change since the previous snapshot. Some repositories install as a bundle, which counts every skill in the bundle. Stars and update dates come from GitHub. Fieldbook is independent of skills.sh and Vercel.

Which version gets installed?

When a source revision is recorded, the install command targets that full repository commit and the source link opens the scanned SKILL.md at the same commit. Older records without a commit remain clearly labelled as unpinned. skills@latest selects the installer version, not the skill version. The SHA-256 identifies only the scanned SKILL.md; referenced files and runtime behavior are not covered by that scan. Human-reading dates and evaluation notes remain separate evidence, not proof that the pinned revision was read in full or tested.

If a refresh fails, a retained review shows the last successful scan for that source. Its pinned command, if present, still targets the last successfully scanned commit. It does not verify newer upstream content. Pins can change after each successful daily scan; save the full command if you need to reproduce an installation. Read the source and its referenced files before installing.

What reviewed does not mean

A review is not a security audit or a guarantee. We read skills but do not run every one, and a skill can change after it is reviewed; the daily checks exist to catch that. Always read what you install. If you find a problem with a listed skill, open an issue (opens in a new tab).