Content
81%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 well-built instruction-only skill body: three clearly separated subcommand workflows, each with numbered steps, explicit validation, and pointed anti-pattern lists, plus disciplined deferral to the AGENTS.md rule layer. The main gaps are the abstract worked example in compact and no concrete pgTAP sample in rls scenarios.
Suggestions
Make the compact worked example concrete: show an actual 3-file local-only chain and the resulting consolidated SQL (even abbreviated), so the merge procedure is copy-paste illustrable rather than purely descriptive.
Include one short executable pgTAP assertion example in rls scenarios (e.g. a positive/negative pair using a seeded persona from `supabase/seed.sql`) so the output shape is unambiguous.
Consider moving the rls persona matrix and silent-edges checklist into a reference file and linking to it from the body, trimming SKILL.md toward the overview role.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and efficient — classification tables, terse anti-pattern lists, no explanation of concepts Claude already knows — with only occasional over-explanation (e.g. the pg_dump aside in the align procedure, 'in your head (or a scratch file)'). It is clearly above the 'mostly efficient but includes some unnecessary explanation' midpoint but not perfectly lean. | 4 / 5 |
Actionability | Guidance is mostly executable: concrete commands (`supabase db reset`, `supabase test db`, `supabase migration new`), a classification table with explicit allowed actions, and specific SQL idioms. It falls short of copy-paste-ready because the compact 'Worked example' stays abstract with no actual SQL, and rls scenarios offers no sample pgTAP assertion. | 4 / 5 |
Workflow Clarity | All three subcommands have clearly numbered sequences with explicit validation checkpoints and feedback loops: user confirmation before destructive merges ('Default to asking', 'Verify the classification with the user when uncertain'), `supabase db reset` smoke checks, and `supabase test db` when pgTAP covers the affected tables. The destructive-operation cap does not apply — validation is present. | 5 / 5 |
Progressive Disclosure | Good structure: clean per-subcommand sections, clearly signaled one-level-deep references to `supabase/AGENTS.md` ('point readers there rather than re-list'), and no nesting. A notch below excellent because the ~250-line single-file body inlines material (the rls persona matrix and silent-edges checklist) that could partly live in a reference file. | 4 / 5 |
Total | 17 / 20 Passed |