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, rigorously engineered workflow: every step is executable, validation and recovery paths are explicit, and disclosure into the references is exemplary. The one real cost is length — the gate and print-in-full rules are restated several times and many sentences are long compounds that could be tightened without losing any rule.
Suggestions
Collapse the three restatements of the print-the-draft rule in step 5 into one directive plus one short brevity-mode carve-out; step 7's copy of the rule only needs to reference it, not restate the rationale.
Split the ~200-word step-1 public bullet into short sub-bullets (fetch order, preflight's draft-capability answer, the scratchpad-subagent rule, comment-only-on-changed-lines) so each instruction is scannable on one line.
Trim the repeated justifications for the verdict rules in step 4 (the 'Holding a PR open...' paragraph largely restates the classify-by-cost rule above it) to a single sentence.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body explains nothing Claude already knows — every paragraph carries novel, skill-specific policy — but it could be tightened. The print-the-draft-in-full requirement is stated three times across step 5's paragraphs ("Print the complete draft as message text", "The draft is the step 4 file, not your memory. Print the file's full content", and the brevity-mode carve-out) and again in step 7; and several sentences are very long compounds (the step-1 public bullet runs ~200 words) that could be split or trimmed without losing rules. This fits the "mostly efficient but could be tightened" anchor, not the "minor instances" of the 4 anchor. | 3 / 5 |
Actionability | Fully executable for an instruction-orchestration skill: exact artifact paths (review/pr-<N>-draft.md, branch names with / replaced by -), exact commands (git remote get-url origin, gh auth status / glab auth status), exact dispatch parameters (subagent_type: "general-purpose", "not fork"), an explicit precedence order for platform resolution, and a hand-off to platform references that contain copy-paste-ready gh/glab commands and JSON payloads (verified in references/github.md). Edge cases branch explicitly (draft unavailable, existing pending review, silent engine, nesting depth limit). | 5 / 5 |
Workflow Clarity | A 7-step numbered workflow with a checklist, explicit validation checkpoints (step 5 AskUserQuestion gate, step 6 verify the summary actually posted, step 7 self-check that the fenced summary visibly rendered), and feedback loops (re-present the revised draft and return to the gate; retry with the comment moved to the summary on out-of-diff rejection; amendments re-enter step 5 before any write). Every failure mode has a named recovery path, matching the top anchor including its error-recovery and checklist criteria. | 5 / 5 |
Progressive Disclosure | The core workflow stays in SKILL.md while platform mechanics (github.md, gitlab.md), the local mode (local.md), engines (analysis.md), voice (house-style.md), and rationale (design-notes.md) are split into six real, one-level-deep reference files, each cited inline at its point of use with its purpose stated. All cited files and sections exist ("Merge and carry forward", "False-positive discipline", "What never goes in a posted review", the re-review examples), and the references themselves carry Contents sections. Structure and navigation match the top anchor. | 5 / 5 |
Total | 18 / 20 Passed |