Guides
Skill Check
Skill Check reviews an AI agent skill and tells you, with evidence, whether it does what it claims to do. It is built for people working with Claude Code, Codex and other coding agents, where a skill folder is easy to write and hard to be sure about.
Nothing in the skill is executed. Skill Check reads the folder, checks the parts it can check in code, and has a model judge only the parts that genuinely need judgement.
What it checks
- Structure, checked in code: the folder layout, the SKILL.md and its front matter, and whether the required pieces are present and well formed.
- Claim against content, judged by a model: does the skill actually contain what its description promises, or does it fall short of it?
- Routing, run as a simulation: given the skill’s trigger description, would an agent reach for it at the right moment and leave it alone at the wrong one?
The routing step is a simulation on the trigger wording, not a live agent session, so it never runs any command the skill describes.
What to upload
Give Skill Check a complete Agent Skill, not a single file. Any of these work:
- A skill folder with its SKILL.md and any supporting files.
- A ZIP of that folder.
- A link to a public GitHub skill.
The more honest the trigger description in your SKILL.md, the more useful the routing result, because that description is exactly what the simulation reads.
How to read the report
The report is evidence first. Each finding tells you what was checked and what was found, so you can act on it rather than take it on trust.
- Structure findings point at the exact file or field, so a fix is usually a quick edit.
- The claim judgement quotes the part of the skill it is reasoning about, so you can see why it agreed or disagreed with the description.
- The routing result shows where the skill would and would not be triggered, which is often where a vague description shows its cracks.
A worked example
A skill that reads well but triggers on the wrong things.
- Skill
- A "commit-helper" skill whose description says it "helps with git", with a clear SKILL.md and a sensible body.
- Structure
- Passed. The folder, SKILL.md and front matter are all present and well formed.
- Routing
- The broad phrase "helps with git" made the simulation reach for the skill on unrelated git questions like resolving a merge conflict, which it does not actually cover.
- Fix
- Narrow the trigger to what it does: "use when writing a commit message for staged changes". The routing simulation then fired only on commit-message requests.
The structure was never the problem. The evidence was in the routing.
Daily allowances
Reviewing a skill runs several model passes, so there is a daily allowance and a shared daily spend ceiling for the whole service. See the usage limits for the detail.
Have a skill you want a second read on?
Open Skill Check