Content
96%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, expert-grade body: every token is project-specific knowledge, commands are executable, and the destructive workflow is gated by confirmation and layered validation with genuine feedback loops. The single soft spot is that the main SKILL.md carries nearly all shared detail inline, making it heavy for an always-loaded overview.
Suggestions
Move the file-by-file Shapes and Backend enumerations (exact .sql filenames, per-shape dependent-file lists) into a shared reference file (e.g., shapes.md / backend.md), keeping only the shape taxonomy and one example per shape in SKILL.md.
Consolidate the thrice-repeated grep-blindness warning (Shapes, Rationalizations, Verification) into the Verification section with a pointer from the others, trimming a few lines without losing the check.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Zero filler or concept explanations — no library tutorials or basic-knowledge padding; every line carries codebase-specific knowledge (grep-invisible nb_<resource>, type-checked i18n catalogs, append-only migrations, per-file SQL responsibilities). Repeated warnings appear across sections, but each repetition serves a distinct phase (discovery, rationalization rebuttal, verification) rather than padding. | 5 / 5 |
Actionability | Copy-paste-ready commands throughout: the script invocation, "make typecheck && make lint", the grep battery, "npx supabase db diff --local -f remove_<resource>", "npx supabase db reset --local", and a sed fallback for the literal-escape Edit failure. Exact file paths (06_grants.sql, root/defaultConfiguration.ts, misc/useImportFromJson.ts) make the editing guidance concrete rather than abstract. | 5 / 5 |
Workflow Clarity | A clear four-step sequence gated by an explicit irreversible-confirmation checkpoint, with strong validation for a destructive operation: typecheck+lint, separate targeted greps, db reset replay, a six-item checklist, and feedback loops ("Resolve whatever tsc surfaces", Edit→sed fallback, db reset catching what db diff misses). The destructive-skill validation cap does not apply — validation is thorough. | 5 / 5 |
Progressive Disclosure | References are one level deep and clearly signaled (per-resource files annotated "spine"/"spine/link", plus cross-skill pointers to backend-dev and writing-migrations). Not a 5: the main file itself is ~97 very dense lines, and the Shapes and Backend enumerations (exact .sql filenames, per-shape file lists) could partially move into reference files to slim the always-loaded surface. | 4 / 5 |
Total | 19 / 20 Passed |