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.
The body delivers exceptionally concrete, executable guidance with strong gating and validation, but it pays for that in token efficiency: heavy verbatim repetition of safety rules, duplicated/malformed step numbering, and a fully inlined monolithic structure with no progressive disclosure. The actionability is near-best-in-class while the conciseness is the clear weak point.
Suggestions
State each safety rule once (e.g., the worktree/no-stash prohibition in the Rules block) and reference it from phases instead of repeating it nearly verbatim four times across Implement steps 2–3, Review 2b, and the CI handoff.
Fix the malformed and duplicated step numbering — Implement's "6. **6. **Hand off to the next phase**" and the Push Phase's step 7 appearing both inside step 6's absent-CI branch and again as a standalone step — so each instruction has exactly one authoritative occurrence.
Trim incident-report-style justification passages (dated measurements, "the usual substitutes…", repeated failure-mode narratives) to the operative rule; this alone would cut a substantial fraction of the ~470 lines without losing executable guidance.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~470-line body restates the same safety rules nearly verbatim in multiple places — the no-stash/no-checkout worktree prohibition appears in Implement steps 2 and 3 ("Never take your uncommitted work out of the working tree… no `git stash` in any form"), again in Review 2b ("Never take the diff out of the working tree"), and a fourth time in the CI handoff ("Do NOT stash, reset, restore or commit") — and the Push Phase duplicates its entire step 7 inside step 6's absent-CI branch before repeating it as step 7. Not 1, because almost none of it explains concepts Claude already knows — it is dense, non-obvious operational guidance — but the duplication and anecdote-style justification passages ("measured 2026-08-19, one run in three…") go well beyond the "some unnecessary explanation" of anchor 3. | 2 / 5 |
Actionability | Fully executable throughout: exact commands (`git rev-parse --abbrev-ref --symbolic-full-name @{u} 2>/dev/null`, `git worktree add <scratch-dir> HEAD`, `git rev-list --left-right --count HEAD...@{u}`), a literal report template with mandatory section order, an ordered first-match-wins verdict-selection procedure, exact user-facing strings to output, and per-stack check-command mappings. Concrete, copy-ready guidance covers all common cases. | 5 / 5 |
Workflow Clarity | The five-phase sequence with four explicit inter-phase gates, a terminal verification checklist, and rich feedback loops (fix-and-rerun on failing tests, baseline-record classification, re-validate-on-error) matches the anchor-5 shape. It drops to 4 because of structural defects that create genuine ambiguity: Implement's step 6 is malformed ("6. **6. **Hand off to the next phase**" with broken nested bold markers) and the Push Phase's step 7 exists twice — once inlined inside step 6's absent-CI parenthetical and once as a standalone step — so an executor must infer which occurrence is authoritative. | 4 / 5 |
Progressive Disclosure | Internal structure is good — distinct phases, gates, a rules block, a verification checklist — but the entire ~470-line workflow is inlined in SKILL.md with no bundle files (references/, scripts/, assets/ do not exist). The per-phase detail (this composite was generated from five separate atom files that were flattened inline) is exactly the content that could live one level deep in per-phase reference files; the no-re-entry rule justifies inlining, but the rubric scores what is written, and the result is a dense single-file wall. Not 2, because what structure exists is clear and consistently organized. | 3 / 5 |
Total | 14 / 20 Passed |