Content
88%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.
An exceptionally actionable, well-validated operational loop: exact commands, explicit clean-state criteria, feedback loops, and stop conditions. The only weaknesses are minor verbosity in steps 6–7 and the Hard Rules recap, and a single-file layout that could offload some detail to a reference file.
Suggestions
Tighten steps 6 and 7 by moving the multi-clause justifications ('A babysit loop spanning a long session is exactly the scenario...') into brief rationale notes, letting the imperative commands lead.
Consider moving the deep /ship sync-check choreography referenced in step 6 into a short reference file (e.g. references/sync-check.md) to reduce SKILL.md inline length while keeping the checklist one level deep.
The 'Hard rules' section duplicates steps 5 and 8; either fold the unique rules (prod-evidence scrubbing, never drop a finding) into the loop steps or keep Hard Rules as the single home and trim the inline repetition.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with workflow-specific gotchas rather than concepts Claude already knows — e.g. the GraphQL pagination guard and the stale-cubic-round caveat earn their tokens. Minor trim candidates exist: the 'Hard rules' section restates steps 5 and 8, and steps 6–7 use long parenthetical sentences that could be tightened. This fits 'efficient; minor instances of over-explanation that could be trimmed' better than the 3 anchor's unnecessary explanation. | 4 / 5 |
Actionability | Fully executable throughout: copy-paste `gh pr view`/`gh api graphql` commands including the complete reviewThreads query, the resolveReviewThread mutation, the reply REST call, exact trigger strings ('@greptile', '@cubic-dev-ai review this PR'), and concrete git commands ('git rebase origin/staging', '--force-with-lease'). Placeholders like <n> and <owner>/<repo> are appropriately parameterized. | 5 / 5 |
Workflow Clarity | A 10-step loop with an explicit 'Definition of clean' requiring three fresh conditions, pending-check handling, pagination until hasNextPage is false, per-round state re-checks, post-push verify, confirm-both-reviewers-pending before waiting, conflict recovery via rebase, and explicit stop conditions. Validation and feedback loops (fix → reply → resolve → re-review, re-trigger if a bot stayed 'pass') are thorough, so the destructive/batch cap at 3 does not apply. | 5 / 5 |
Progressive Disclosure | No bundle files exist, and the single file is well-sectioned (When to use, Inputs, Two reviewers, Definition of clean, Loop, Reporting, Public-repo hygiene, Hard rules) with one-level cross-references to the /ship skill for delegated detail. At ~190 lines with some inline operational detail (e.g. the sync-check narrative in step 6) that could live in a reference file, this sits at 'good structure; minor organization gaps' rather than the fully split 5 anchor. | 4 / 5 |
Total | 18 / 20 Passed |