Content
85%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A dense, highly executable orchestration skill with excellent validation checkpoints, explicit error-recovery paths, and proper use of its two bundle assets. The main weakness is systematic internal repetition: the auto-defer, reviewer, and comment-posting rules are each restated two to three times across the Activation card, Origin and writes, Procedure, and Hard nos sections.
Suggestions
State the auto-defer condition once (e.g., in 'Origin and writes') and have the Activation card and 'Hard nos' sections reference it, instead of restating the full three-line condition in each place.
Trim 'Hard nos' to only rules not already stated earlier (merge/push and running sibling workers), removing the recaps of the reviewer, deferred-label, and label-write rules.
Consolidate the duplicated autopilot-comment posting guidance (debug/json/public_body handling appears in both the Activation card rules and Procedure step 6) into one authoritative block.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is telegraphic with zero concept padding, but the same constraints are restated systematically: the auto-defer condition appears three times ('Write `status/deferred` only when this PR is not labelled `status/accepted`...' in Activation card; 'Auto-defer: if this PR is not labelled `status/accepted`...' in Origin and writes; 'Do not write `status/deferred` when a linked issue is `status/accepted`...' in Hard nos), and 'Never assign. Never request reviewers' and 'Do not call `gh pr comment` yourself' each appear 2-3 times, with Hard nos largely recapping earlier rules. Not 4 because the redundancy is systematic rather than minor; not 2 because every line carries operational content with no filler or explanation of concepts Claude already knows. | 3 / 5 |
Actionability | Concrete and executable throughout: copy-paste Enter/Exit card templates with every field enumerated, exact label vocabulary ('ready-for-review | needs-design | needs-issue | duplicate-of | decline-with-reason | auto-handle'), exact paths to real files (assets/label-contract.json, assets/pr-triage-template.md), the exact CONTRIBUTING.md URL, and explicit defaults ('`write` defaults to `on` when the caller omitted it'). The common cases (accepted linked issue -> ready-for-review; no accepted issue -> needs-issue + status/deferred + thank-the-author) are concretely covered. Not 4: no material execution gaps for an instruction-only skill. | 5 / 5 |
Workflow Clarity | The six-step Procedure has explicit stop checkpoints at every risky boundary ('Missing number -> stop', 'Missing card -> stop', autopilot-comment probe 'Missing -> stop. Do not post'), an error-recovery feedback loop ('Missing footer after that write is a failed post -- patch via `autopilot-comment` (`intent: patch`)'), and a dedup no-op ('If a prior triage comment's watermark still matches HEAD conversation, no-op'). Not 4: checkpoints are explicit, not implicit, and the operation is single-PR advisory so the destructive/batch cap does not apply. | 5 / 5 |
Progressive Disclosure | Two real one-level-deep bundle files, both clearly signaled and verified present on disk: 'Read `assets/label-contract.json` (same contract as issue triage) before reasoning' and 'Fill [assets/pr-triage-template.md](assets/pr-triage-template.md)'. The bulky label taxonomy and the output template are correctly externalized while orchestration logic stays in SKILL.md. Not 4: the references are explicit, present, and only one level deep. | 5 / 5 |
Total | 18 / 20 Passed |