Content
73%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 strong operational runbook with an exceptionally clear step sequence, explicit STOP conditions, and real feedback loops — the unattended workflow design is unambiguous. Its weaknesses are redundancy (the one-spawn and never-self-approve rules each appear several times, and Common Mistakes overlaps the Red flags table) and a reviewer-subagent dispatch that describes rather than scripts the gate.
Suggestions
Merge the Common Mistakes and Red flags sections into one — they cover the same failure modes (asking when the codebase answers, self-approving tests, letting an intermediate phase spawn) in slightly different phrasings, saving ~15 lines.
State the one-spawn rule once in the Core principle and reference it from Steps 3, 4, and 6 instead of restating the rationale at each step.
Provide the Step 5 reviewer subagent's prompt verbatim (or a template) so the approval gate is dispatched exactly the same way on every run rather than assembled from a description.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is operational and free of explanations of known concepts, but the one-spawn rule is restated in the intro, core principle, Steps 3/4/6, Common Mistakes, Red flags, and Important, and the Common Mistakes and Red flags sections largely duplicate each other ("Stopping to ask when the codebase already answers it" vs "The ACs are clear but I'll ask anyway to be safe"). | 3 / 5 |
Actionability | Mostly executable: the worktree creation bash block is copy-paste ready, the ticket-key regex and exact phase-handoff invocations are concrete. Minor gaps remain — the Step 5 reviewer subagent dispatch is described but no ready-to-use prompt is given, and placeholders like `<short-kebab-summary>` and `<resolved main clone>` are underspecified. | 4 / 5 |
Workflow Clarity | Steps 0–6 are clearly sequenced with a mermaid overview and explicit validation checkpoints throughout: branch confirmation with a STOP on mismatch, the ambiguity decision tree with human-only forks as blockers, the reviewer gate with a NEEDS_WORK fix-and-re-dispatch loop and a two-strike stop, and worktree-creation failure handling. | 5 / 5 |
Progressive Disclosure | A single-file skill with no bundle references, but the content is a coherent operational runbook organized into well-labeled numbered steps and sections that are easy to navigate. At ~139 lines with everything inline and some cross-section redundancy, structure is good rather than optimal. | 4 / 5 |
Total | 16 / 20 Passed |