Content
82%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 strong, high-density reference skill: executable SQL examples, exact dialect/limit constraints, denied-table guidance with SDK alternatives, and a clear tool-routing table. The main improvement opportunities are trimming verbose SDK-table rows and splitting the schema catalog into a one-level-deep reference file.
Suggestions
Trim the behavioral parentheticals in the host.delegate and host.submit_output rows of the SDK table (backgrounding, interrupt semantics, bounce handling) into a short separate note or the relevant skill, keeping the table rows scannable.
Move the per-table column listings (frames, artifact_versions, execution_log, etc.) into a references/schema.md and keep SKILL.md to one-line table summaries plus the worked examples, so the overview file stays lean.
Add a brief explicit error-recovery note for the two documented failure modes (truncation → narrow columns / aggregate server-side; timeout → simplify the query) as a short checklist, making the feedback loops as actionable as the query guidance.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense reference material that assumes Claude's competence — no SQLite tutorials, no basic-concept padding — with high information per line ('json_extract(col, $.key)', epoch-ms examples). A few spots could be trimmed, e.g. the long behavioral parentheticals in the host.delegate and host.submit_output SDK rows and the repeated 'only available via the repl tool' note. | 4 / 5 |
Actionability | Four complete, copy-paste-ready worked examples ('SELECT COUNT(*) … SUM(total_cost) FROM frames', execution_log with json_array_length(e.files_written), artifacts JOIN artifact_versions) plus concrete introspection commands ('PRAGMA table_info(frames)', sqlite_master query) and exact limits (limit=1000, 2000-char cell cap, 5s timeout). | 5 / 5 |
Workflow Clarity | This is a reference skill whose single action is unambiguous: run read-only SQL via the repl tool, with a clear sequence (dialect/limits → schema introspection → per-table queries) and built-in error-avoidance feedback ('truncated=True … narrow your columns', 'Aggregate server-side so the row cap can't undercount', prefer host.frames() over raw _messages). It stops short of an explicit validate-and-retry loop, so it fits 'clear sequence with most checkpoints present' rather than the full-checklist anchor. | 4 / 5 |
Progressive Disclosure | The single file is well-sectioned ('Dialect and limits', 'Queryable tables' with per-domain subsections, 'Denied tables', 'Worked examples', 'SDK surface') and its only external pointers are to other skills ('load the customize skill') at one level deep. However, the ~150-line table/column catalog is all inline in SKILL.md; a references/schema.md split for the per-table column listings would move it to the top anchor. | 4 / 5 |
Total | 17 / 20 Passed |