Content
71%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.
The body is an admirably lean orchestration adapter with an explicit lane sequence and a hard validation gate, and it wastes no tokens. Its main weakness is that the canonical spec it depends on — pr-swarm-review/orchestration.md and the personas directory — is not present in the provided bundle, leaving the bulk of the executable detail unreachable from this file.
Suggestions
Ship the referenced canonical spec inside the skill bundle (e.g. under references/ or as pr-swarm-review/orchestration.md) so 'read it now and follow it' resolves to a real file, or inline the minimal lane definitions needed to run a review if the spec must stay external.
Include the persona filenames or lane names inline (even just a one-line list of lanes 1–7) so a reader can verify the swarm structure without the external file.
Add a fallback instruction for when orchestration.md or the gitnexus-* project agents are missing, so the skill fails gracefully instead of dead-ending.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~30-line body is lean and imperative with zero padding: 'Run in Swarm mode. Dispatch each lane as its own subagent via the Agent tool.' It assumes Claude's competence throughout — no explanation of what a PR review is, what a subagent is, or why orchestration matters. It matches the anchor 'lean and efficient; assumes Claude's competence; every token earns its place'; even the blockquote distinguishing the CI skill carries real, non-redundant information. | 5 / 5 |
Actionability | There is genuinely concrete guidance — the invocation '/gitnexus-pr-swarm-review <PR URL or PR number>', 'the seven subagents are the project agents named gitnexus-*', 'Run lanes 1–2 first, lanes 3–6 in parallel after, and lane 7 last' — but the operative content (the full review contract: lanes, classifications, finding format, hidden-Unicode checks) lives in 'pr-swarm-review/orchestration.md', which is not present in this skill's bundle (no references/, scripts/, or assets/ directories exist, and no pr-swarm-review/ directory is shipped). So the key execution details are unreachable from what is provided, matching the anchor 'some concrete guidance but incomplete; missing key details'. It is not score 4 because the missing spec is the majority of the executable guidance, not a minor gap. | 3 / 5 |
Workflow Clarity | The sequence is explicit — 'Run lanes 1–2 first, lanes 3–6 in parallel after, and lane 7 last on the draft' — and it includes a real validation checkpoint with a feedback loop: 'Lane 7 is a hard gate. Do not emit the final review while the synthesis critic's "Required corrections before posting" section is non-empty — revise and re-run it.' This matches anchor 4 ('clear sequence with most checkpoints present; minor validation gaps'). It is not score 5 because the per-lane steps themselves are delegated to the absent orchestration.md, so the checkpoints within lanes 1–6 are not visible here; it is above anchor 3 because an explicit gate-and-retry loop is stated rather than merely implied. | 4 / 5 |
Progressive Disclosure | The split is conceptually right — an adapter that pins Claude Code specifics while pointing to a canonical spec ('pr-swarm-review/orchestration.md — read it now and follow it', personas under 'pr-swarm-review/personas/') — and the references are clearly signaled and one level deep. However, no bundle files are provided: the referenced orchestration.md and personas/ directory do not exist in this skill's directory, so navigation dead-ends and the disclosure structure cannot actually be followed. That lands between anchor 3 ('some structure but could be better organized; references present but...') and anchor 4; the clearly signaled but unresolved references keep it from scoring 4. | 3 / 5 |
Total | 15 / 20 Passed |