A Quarter of Skills State What They Won't Do — Just Not Where It Matters
The negative space is not missing. It is in the wrong place.
Two counts across the same 765 skills:
| Where the boundary appears | Skills | Share |
|---|---|---|
| In the SKILL.md body | 196 | 25.6% |
| In the one-line tagline | 4 | 0.5% |
A quarter of authors have thought about what their skill should not do. Almost none of them put that thought where it does the most work. We previously found that only 0.5% of taglines state a boundary; this post asks where the other boundaries went.
Why the placement matters
The description is the routing table. An agent reads it, decides whether the skill matches, and only then loads the body. A boundary that lives in ## Boundaries or in the sixth paragraph of ## When to Use is read after the decision to load has already been made.
So the 196 authors wrote the right sentence and put it in the wrong file section. The skill will still fire on adjacent requests; the exclusion only helps once the agent is already committed.
What the body versions look like
The body boundaries are usually good sentences. A representative shape:
> This skill does not cover drafting replacement language, filing, signing, or sending anything.
That is exactly the kind of statement that belongs in the description. It names adjacent requests and tells the router to look elsewhere.
The arithmetic of moving it
Moving a boundary from the body to the description costs nothing — it is the same sentence, relocated. It changes the skill from "fires, then explains its limits" to "does not fire".
For an author, that means:
- Open the frontmatter.
- Take the first paragraph of
## Boundaries, or the "does not" sentence from## When to Use. - Compress it to one clause:
Not for X, Y, or Z. - Append it to the description.
The 25.6% who already wrote the sentence are one edit away from the behaviour they intended.
The caveat
We measured the association, not the effect: body boundaries correlate with a modest score gap (+0.34), well within our evaluator's noise. What we are confident about is the mechanism, not the size — a description-level exclusion changes routing behaviour in a way a body-level one cannot.
The 0.5% figure for taglines comes from the same tagline corpus we analysed separately. Both numbers describe the same directory on the same day.