Content
81%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 and well-validated orchestration skill: exact commands, formats, checkpointed steps, and mature failure handling. Its weakness is density — rhetorical justification prose and content duplicated with the reference file make the body longer than the instructions alone require.
Suggestions
Trim the multi-sentence rationale paragraphs (e.g., step 3.2's 'The split is the point…' and the split-angle prose in step 2) to one-line rules with the reason attached — the same points are made repeatedly across sections.
Keep the confidence/severity tables and False Positive Examples only in references/subagent-prompts.md (where they already exist as verbatim shared blocks) and have the body point to them, rather than duplicating both inline.
Move the detailed gh Command Reference recipes (e.g., the past-PR-feedback xargs pipeline used only by agent #4) into the reference file next to the agent templates that consume them.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body avoids explaining concepts Claude already knows and nearly every paragraph carries procedural load, but the prose is expansive — multi-sentence rationales (e.g., "The split is the point. A defect is usually a place where changed code at one location relies on a fact at another location…") and repeated justifications for the same rule could be tightened without losing instruction value, matching 'mostly efficient but could be tightened' rather than the lean score-5 anchor. | 3 / 5 |
Actionability | Guidance is fully executable throughout: copy-paste `gh` commands with exact `--jq` filters ("gh api repos/OWNER/REPO/pulls/78 --jq '.head.sha'"), a worktree setup/teardown script with collision-safe naming, a complete batched-review JSON payload, and exact link and comment formats covering the common posting cases. | 5 / 5 |
Workflow Clarity | A clearly sequenced 9-step workflow with a mandated todo checklist, validation checkpoints at every gate (eligibility check in step 1, re-check in step 6, adversarial verification in step 4, the confidence/severity filter in step 5), explicit error-recovery feedback loops (retry-once rules, unscored vs confidence-0 distinction, double-post avoidance for `gh` failures), and unconditional teardown — matching the score-5 anchor including its feedback-loop and checklist criteria. | 5 / 5 |
Progressive Disclosure | The subagent prompts are appropriately split into the real, one-level-deep, clearly-signaled references/subagent-prompts.md, and the body is well-sectioned; however, the confidence/severity tables and the False Positive Examples appear both inline in the body and as shared blocks in the reference file, and the body itself runs ~370 lines, so some content still lives inline that could rest solely in the reference — 'good structure, minor organization gaps' rather than a clean score-5 split. | 4 / 5 |
Total | 17 / 20 Passed |