| name | insight-rate |
| description | Use this skill when the user (or another Claude instance) wants to rate a previously injected insight as good / bad / irrelevant for the current task. Triggered by "/insight-rate", "rate this insight", "this lesson saved me time", "this lesson was wrong", "thumbs up the insight", or similar phrasings. Writes a single rating record to the per-repo ratings.jsonl. Used at Gate 1 to compute "≥1 cross-engineer relevant retrieval / week" and detect drift in topic-tag inference. |
| argument-hint | <lesson-id> <good|bad|irrelevant> [reason] |
| allowed-tools | Bash, Read |
/insight-rate
Tag an injected lesson as good, bad, or irrelevant.
The rating is the only Gate-1 metric that requires explicit user action — keep
it lightweight. The rating is appended to the local ratings.jsonl and
synced to the team mirror on the next push, so all collaborators see the
verdict.
Required arguments
/insight-rate <lesson-id> <good|bad|irrelevant> [reason]
lesson-id: the id printed in the <insights-share> injection block.- verdict: one of
good / bad / irrelevant.
reason (optional): a short free-text note (saved verbatim, untouched by
the Layer-2 PII filter — keep it generic).
Implementation
The skill is implemented by scripts/rate-lesson.sh. The script:
- Resolves the canonical repo slug via
canonical-remote.sh.
- Appends
{lesson_id, verdict, reason, rated_by, rated_at, repo_slug} to
~/.gstack/insights/<repo-slug>/ratings.jsonl.
- Schedules a background
sync-mirror.sh push so the rating reaches the
mirror without blocking.
Why we keep it manual
The design doc explicitly suppresses an inline thumb prompt in v1 — the
plugin must stay user-unaware during normal use. /insight-rate is invoked
only when the receiving engineer chooses to rate; routing logic does not
depend on it. (Gate 1 metrics depend on it.)
Examples
User: /insight-rate lesson-1714044020-abc12345 good
> wrote rating to ~/.gstack/insights/<slug>/ratings.jsonl
User: /insight-rate lesson-... bad — wrong stack
> wrote rating + reason
