Content
67%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.
A dense, opinionated, well-structured process skill: real commands, ready-to-use templates, a clear sequenced workflow with validation checkpoints, and an exemplary set of one-level-deep worked examples covering all five feature shapes. Its main weaknesses are repetition of the same key ideas across sections and heavy inlining of detail that pushes the main file well past lean.
Suggestions
Consolidate the repeated registry-driven/differential guidance into one canonical section (e.g. the 'differential gate' axis) and cross-reference it from step 1, step 4, and the migration shape instead of restating it — this would lift conciseness without losing content.
Move the 'Common axes worth probing' catalog (~40 lines) and the per-shape strategy detail from step 1 into a reference file, keeping one summary line per shape in the main body, so SKILL.md reads as the overview the rubric expects.
Link each worked example inline at the shape it demonstrates in step 1 (e.g. '*port* — see references/example-identity-resolution.md') so the references are discoverable at their point of use, not only at the end of the document.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is almost entirely non-obvious, experience-distilled guidance rather than concepts Claude already knows, but key ideas repeat across sections: registry-driven harnesses are explained in step 1, step 4, the migration-platform shape, the 'Reusable per-wave gate' axis, and the acceptance template; differential exact/known-diff/merge tagging appears in step 5, the 'differential gate' axis, and the template. Not 4 (repetition goes beyond 'minor instances that could be trimmed'); not 2 (none of it is generic padding). | 3 / 5 |
Actionability | Concrete, copy-paste-ready artifacts are present — 'gh issue view <n> --repo constructorfabric/insight --json title,body,labels,assignees,projectItems', 'git worktree list --porcelain', and two full output templates — plus verb-led directives ('Build the metric×view harness…'). Not 5: a fair share of guidance stays abstract ('Trace where the thing under test actually reads vs writes'), and the concrete per-case detail lives mostly in the external reference files rather than inline. | 4 / 5 |
Workflow Clarity | A clear numbered 1–8 sequence with genuine checkpoints — 'resolve specification disagreements before claiming a test group's coverage', 'report the mismatch before changing expectations', 'confirm the boundary with the user rather than silently excluding', 'Draft for review first; don't auto-create' — and an acceptance checklist in the template. Not 5: the density of interleaved principles inside each step and the '1b' insertion make the underlying sequence harder to follow than the anchor-5 exemplar, and some checkpoints are embedded in prose rather than framed as validate-then-proceed steps. | 4 / 5 |
Progressive Disclosure | The five worked examples are properly split into references/ (all verified to exist), one level deep, each clearly labeled with the shape it models and what it demonstrates. Not 5: substantial detail remains inlined in SKILL.md (the ~40-line 'Common axes' catalog, per-shape strategy in step 1) that could live in references, and the example files are only linked at the end rather than signposted at their point of use in step 1. | 4 / 5 |
Total | 15 / 20 Passed |