Reading Your Discovery Gaps

Most analytics tell you what worked. The discovery-gaps view tells you what didn't — the searches where an agent went looking and your registry came up empty. It's the most valuable list you have, because real usage wrote it for you.

What a gap is

Every agent search runs against your skills semantically, returning a best similarity score (similarity = 1 − distance). A gap is a search whose best match scored low — nothing in your registry came close enough to be useful. The agent asked; you had no good answer.

QUERY                                     BEST MATCH                 SCORE   COUNT
"redact PII from a pdf"                   Extract tables from PDFs    0.41     38
"summarize a slack thread"               (none above threshold)      0.22     27
"convert markdown to a confluence page"   Publish docs to Notion      0.49     19
"generate an OpenAPI spec from code"     (none above threshold)      0.18     14

Each row is a request agents kept making that you never served well.

Why it's your most valuable list

A roadmap written from real demand beats one written from guesses. The gaps view is exactly that: a backlog ranked by how often agents needed something you couldn't give them. High count plus low score means recurring, unmet demand — that's where the next skill pays off fastest.

This is the heart of the query is the telemetry: you don't survey users about what skills to build, you read what their agents already tried to do.

Triaging a gap

Not every gap means "write a new skill." Read the best match and its score:

• Low score, no relevant skill exists → genuine coverage hole. Author a new skill. "summarize a slack thread" at 0.22 with nothing close is a clear new-skill candidate.
• Low-ish score, but a relevant skill does exist → discovery problem, not a coverage problem. The skill is there but its trigger phrases don't cover this phrasing. "redact PII from a pdf" matching your PDF skill at only 0.41 means: add that phrase to the existing skill.
• High count, any score → prioritize it. Frequency is demand.

From gap to fix

Every gap row links straight back into the editor. Click through, and you either open the near-miss skill to strengthen its description and trigger phrases, or start a fresh skill from the query itself. Because skills are immutable and versioned, your fix ships as a new version without disturbing what already works.

Then watch the same query reappear next week — at a score that finally clears the bar. For how the analytics fit together, see the analytics overview.