Content
62%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 ship workflow itself is a model of clarity — precisely sequenced steps, explicit gates, and genuine feedback loops — and almost every instruction is executable. The score is dragged down by severe token bloat: hundreds of lines of framework boilerplate, persona/voice rules, and over-explained audit methodology inlined into a 1,755-line monolith, with all reference files living outside the bundle.
Suggestions
Move the ~340 lines of pre-Step-0 framework boilerplate (Voice/persona rules, AskUserQuestion Format, Completeness Principle, Contributor Mode, telemetry flows, Plan Status Footer) into a shared reference file cited once, so SKILL.md starts at the ship workflow itself.
Extract the Step 3.4 coverage-audit methodology (data-flow tracing, user-flow mapping, ASCII diagram conventions) into a reference file, keeping only the gate thresholds, regression rule, and output format inline.
Remove hardcoded project-specific commands ("bin/test-lane", "npm run test", the eval-suite glob patterns for one Rails codebase) and derive the test command from the framework detection in Steps 2.5/3.4 so the skill is executable on arbitrary repos.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Roughly a third of the 1,755-line body is padding: ~340 lines of framework boilerplate before Step 0 (Voice persona with writing rules like "No em dashes", AskUserQuestion format, Boil-the-Lake effort tables, telemetry and Contributor Mode flows), "100% test coverage is the key to great vibe coding... it's a superpower" philosophy, and a ~230-line coverage audit that over-explains concepts Claude already knows ("Where does input come from? (request params, props, database, API call)", "Double-click/rapid resubmit"). This matches anchor 2 ('noticeably verbose; several unnecessary explanations or padded sections') rather than 3, whose 'mostly efficient' standard is not met at this volume; not 1 because the core workflow steps are genuinely operative rather than basic-concept tutorials. | 2 / 5 |
Actionability | Nearly every step ships concrete, executable commands ("git fetch origin <base> && git merge origin/<base> --no-edit", "gh pr create --base <base> --title ...", exact AskUserQuestion option blocks, exact commit-message and JSONL log formats). Score 4 rather than 5 because Step 3 hardcodes "bin/test-lane" and "npm run test" for one specific repo (contradicting the framework-agnostic bootstrap in Step 2.5), and placeholders like "{detected test command}" are left unresolved; not 3 since the overwhelming majority of the guidance is concrete and executable. | 4 / 5 |
Workflow Clarity | Steps are numbered 0 through 8.75 with explicit stop conditions enumerated up front, validation gates throughout (Step 3 test gate, Step 3.4 coverage thresholds, Step 3.45 plan-completion audit, Step 6.5 "IRON LAW: NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE"), and explicit feedback loops ("Fails → fix once. Still fails → delete silently"; "If tests fail here: STOP... return to Step 3"; re-run /ship after review fixes). This is an exact match for anchor 5. | 5 / 5 |
Progressive Disclosure | No reference files exist in the skill's bundle; referenced paths ("$GSTACK_ROOT/ETHOS.md", ".agents/skills/gstack/review/checklist.md", "review/greptile-triage.md", "../qa-only/SKILL.md") point to an externally installed framework, while ~570 lines of shared boilerplate and coverage-audit methodology are inlined monolithically. Anchor 3 fits ('content that should be separate is inline'); not 4 because 'most content is appropriately placed' fails when a third of the file is misplaced boilerplate, and not 2 because section structure exists and references are clearly signaled rather than buried. | 3 / 5 |
Total | 14 / 20 Passed |