Content
63%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 extremely thorough, highly actionable orchestration skill: every command, parameter, template and failure branch is concrete and executable, with real validation and retry loops around its risky operations. Its weaknesses are token efficiency — massive run-on paragraphs with substantial cross-section redundancy — and the absence of any progressive disclosure, with all reference material inlined in a single ~13k-token file.
Suggestions
Move stable, bulky material into one-level-deep bundle files — e.g. references/brief-template.md and references/helper-refusals.md — so SKILL.md carries only the workflow, cutting the always-loaded context substantially.
Break the 400-word nested steps (Start steps 3, 5, 6; Stop step 2) into short bulleted sub-steps or a table of refusal → action, which would tighten conciseness and make the sequence easier to follow.
State each cross-cutting rule once (checkedRecently/worktreeWork caveat, who-drives-the-watch decision tree) in a single section and reference it from Start, Round, Status and Resume instead of restating it in each.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is a ~54KB wall of run-on prose — single steps span 400+ words of nested parentheticals (e.g. Start step 3 and step 5), and the same logic is restated across sections: the checkedRecently/worktreeWork caveat appears in Start steps 3 and 5, Resume step 3 and Status, and the who-drives-the-watch decision tree is re-derived in Round, Stop, Status and Resume. Nothing explains concepts Claude already knows, but the cross-section repetition and nesting are noticeably verbose and could be consolidated into a fraction of the tokens. | 2 / 5 |
Actionability | Fully executable throughout: exact helper commands ("${QWEN_CODE_CLI:-qwen}" autofix-pr prepare <n> --repo <owner>/<repo> --host <host>), exact tool-parameter mappings (--take-over → take_over: true, --no-post → no_post: true), exact file paths, a complete fill-in brief template and a complete JSON plan schema with worked values. Copy-paste-ready guidance covers the common cases, matching the anchor-5 example. | 5 / 5 |
Workflow Clarity | Sequences are explicitly numbered (Start 1–9, Round 1–9, Stop 1–3) with exhaustive refusal/answer branching, retry limits ("up to three times in a round"), verification-before-commit in the brief, and rollback/discard rules around push — genuine feedback loops for the risky operations. It falls short of 5 only because single steps embed 400-word nested conditionals whose ordering is hard to follow in edge cases, though no validation step is actually missing. | 4 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/ or assets/), so everything is inlined in one monolithic SKILL.md. Sections are well-labeled and navigable, but content that clearly belongs in separate files — the brief template, the per-helper refusal catalogs, the Status decision tree — is inlined, inflating the always-loaded context. Anchor 3 fits: some structure, but content that should be separate is inline. | 3 / 5 |
Total | 14 / 20 Passed |