Content
92%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 exemplary phase-based playbook: concrete commands, explicit validation gates, and a well-structured one-level reference bundle that matches the files actually present. The only weakness is minor verbosity in the Step 3.3 policy prose and some duplication between the Phase 2 and Phase 5 fixup/rebase sections.
Suggestions
Rewrite Step 3.3's behavioral-impact-review paragraphs as 2–3 terse bullets and move the resolution/fallback rules into a reference file — the hedging prose ('This does not invoke an independent review or grant maintainer authority...') is the densest non-actionable text in the skill.
De-duplicate the fixup/autosquash workflow: Phase 2.3 and Step 5.2 both describe it — keep one canonical inline treatment and make the other a pointer to references/phase2_implementation.md.
Clarify or delete the 'never stop another session's service to make a test pass' clause in Step 3.1 — as written it is cryptic and reads as boilerplate rather than instruction.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly lean, command-driven, and assumes Claude's competence (no padding explaining what PRs or CI are), but a few passages don't earn their place: Step 3.3's dense policy hedging ("Resolve that reference through the installed github-review-pr Skill when contributing outside this marketplace... This does not invoke an independent review or grant maintainer authority"), Phase 3.1's odd "never stop another session's service to make a test pass" clause, and the fixup/autosquash workflow appearing in both Phase 2.3 and Step 5.2. Not 5: those trimmable spots exist; not 3: everything else is concrete and tightly written. | 4 / 5 |
Actionability | Fully executable throughout — copy-paste-ready commands with real flags (gh pr list --jq filter, gh api ... -F in_reply_to for bot replies, git -c sequence.editor=: rebase -i --autosquash, --force-with-lease), a complete PR-body skeleton, a disclosure template, and a worked test-coverage matrix. Specific examples cover the common cases (baseline sizing, conflict rebase, screenshot placeholders). | 5 / 5 |
Workflow Clarity | Phases 0–5 are clearly sequenced with explicit validation checkpoints and feedback loops: the Step 3.3 "What's my evidence?" self-audit, the Step 3.4 push-time checklist (visibility, mergeability, no --no-verify), conflict-resolution recovery, and --force-with-lease semantics for the destructive force-push step. Anchor 5's checklist-and-recovery pattern is met; the destructive-operation validation cap does not apply because validation is present. | 5 / 5 |
Progressive Disclosure | A clear overview playbook pointing to 12 one-level-deep reference files, each inline-signaled ("See references/X for...") and indexed in a Reference Files table with per-file purpose; all in-bundle referenced paths resolve against the actual references/ listing. The one out-of-bundle path (../github-review-pr/references/behavioral-impact-review.md) is explicitly scoped with fallback behavior for when it is unavailable, so navigation is never broken. | 5 / 5 |
Total | 19 / 20 Passed |