Trust · rules last changed 22 October 2026

How we check

A skill is text an AI obeys. Before one is shared with the team or sent outside, the check reads every file as text, and Claude reads it for what rules cannot see. It never runs the content. It says “checked”, with a date. It never says “safe”.

IT MEANS

IT DOES NOT MEAN

Every file of this exact version was read as text, by the rules and by Claude, on this date

That the skill is safe in every situation

A person decides on what the check flags

That nothing can be found. Rules can be fooled, and so can a model that reads

It was checked on its own

That it is harmless in combination with other skills

A finding is a reason to look

That a finding proves bad intent

What the check reads

Form and limits

Names, sizes and kinds of file. Every file is read as text, or the skill is refused.

Patterns and hidden text

Comments, invisible characters, styled-away elements, even inside an SVG logo. Encoded blocks, links that carry data, commands that fetch and run, keys.

Links and packages

Every link and package the skill names, against the lists of known bad ones.

Text aimed at the model

Instructions written for the AI rather than for the reader. A quick second opinion, passage by passage.

A reading by Claude

Does it do what its description says, and nothing else? Two passes, a third when the two disagree; every finding names its line.

A person

Decides on what the check flags. A finding is a reason to look, not a verdict.

When it runs

Advice before sharing. Shown to the member as they share a skill with the team, on the spot, line by line: anything that usually should not leave their machine. The lead sees the same notes.

The full check before a copy leaves. Before a skill is sent to someone outside the workspace, every layer runs, in fifteen to twenty seconds. The copy carries the date of its check.

Before a publish. Every publish is validated with Claude’s own plugin validator before it goes out, so what reaches everyone’s Claude is what the lead approved.

What comes of it

pass

Nothing was found. The skill is checked, with the date, and goes where it was going.

notes

Something to look at, line by line. The person sharing or sending sees it on the spot; a lead sees the same notes. They decide.

refused

A rule was broken, or the check could not finish. Nothing leaves on an unfinished check.

A skill that passes is checked, with a date. The next version is checked again. When the rules change, the date tells you which rules read it.

The limits, said plainly

Rules can be fooled, and so can a model that reads. The check reads a skill on its own, not in combination with the other skills a person has, and not with their files, connectors or permissions. It reads text; it does not run anything. That is why a person decides on what it flags, and why no tool should promise more.

Report a problem

Something in a skill that the check should have seen, or a checked skill that behaves badly: write to hello@skillkeeper.ai, or use Report a problem from the account menu in the app. Line numbers help. Reports are read within a day. What SkillKeeper keeps and who sees it is on the Security page.