Skill Manifest Check · comparison
A careful reviewer genuinely catches a missing field or an obviously short description. This page names no company or product — reading a SKILL.md before publishing it is a habit most skill authors already have, not a competing tool.
A missing name or description field, an obviously malformed frontmatter block, a description that is a few words long — those are visible on the page, and a reviewer who knows to look for them catches them the same way this tool's frontmatter checks do.
| Case | Reading by hand | This tool |
|---|---|---|
| A description with no trigger cue | Can read as clear, grammatical English while giving a dispatcher nothing concrete to match a real request against -- "clear" and "functions as a trigger" are different judgments | Checks for an explicit "use when...", an example phrasing, or a quoted sample request, separately from whether the prose reads well |
| Two skills competing for the same request | The defect lives in the relationship between two files, not in either one -- requires holding every other skill's exact wording in mind while reading each one | Runs a distinctive-vocabulary overlap check across every pasted skill at once, the same way on every run |
| An oversized instruction body | Looks short as rendered on screen; token count and screen length are not the same measurement | Estimates tokens from character count and flags against a documented caution line |
| A description that is long enough but still generic | Word count is easy to eyeball; whether the words used are actually distinctive to this skill's job is a slower, easier-to-skip judgment | Checks length and distinctive vocabulary as separate signals, after removing a fixed filler-word list |
Whether a skill's description competes with a sibling's is not something the file itself can tell you by being read carefully — it depends on what every other skill in the same bundle says, and that comparison gets harder to hold in your head as the bundle grows. A mechanical Jaccard overlap check over the distinctive words in each description does not get tired or skip a pair; it runs the same comparison across all of them every time.
This is a structural, pattern-matching scan, not a real trigger test inside an agent runtime. It cannot tell you whether a specific description will actually fire on a specific request, and a skill that trips no finding here can still fail to trigger in practice. Judging whether the skill's instructions are actually correct for the task, and whatever wording this tool's heuristics don't recognise, is still a job for reading the file yourself.
A reviewer catches a missing frontmatter field, an obviously malformed name, or a description that is simply too short. Those are real defects and a careful read finds them the same way this tool's frontmatter checks do.
A description can read as clear, well-written English to a human and still give an agent's dispatcher nothing concrete to match a real request against. Judging whether a sentence actually functions as a trigger cue -- an explicit "use when...", an example phrasing, a quoted sample request -- is a different read than judging whether the sentence is grammatical, and the two can disagree.
Because the defect isn't in either file -- it's in the relationship between them. Two descriptions can each be well-written on their own and still share enough distinctive vocabulary that a dispatcher choosing between skills for one request could reasonably pick either. Reading skill A and later reading skill B does not surface that unless the reviewer is holding every other skill's exact wording in mind at the same time, across however many skills the bundle has.
A skill's instruction body can look short on screen and still run long in tokens once headings, code fences, and examples are counted -- length as rendered and length as consumed by the model are not the same measurement, and a reviewer scrolling the file is judging the first one.
No. Every check here is a structural, pattern-matching signal against known failure shapes, not a real trigger test inside an agent runtime. A skill with zero findings can still fail to trigger in practice, and this tool has no way to observe that -- read a clean result as "no known authoring mistake found," not as a guarantee.