Content
71%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 body is a dense, highly actionable operational guide: every command and SQL snippet is executable, decision rules for preview/run/deploy are explicit, and validation checkpoints (dry-run, diff, preview-before-deploy) are woven in. Its weaknesses are repetition of the same deploy/metadata warnings three to four times, and a single-file layout that inlines general Windmill script-writing guidance into a PostgreSQL-specific skill with inconsistent heading levels.
Suggestions
Deduplicate: state the 'deploy only when explicitly asked' and 'generate-metadata is not a deploy' policies once in a single policy block instead of repeating them in the command list, the preview-vs-run section, and the after-writing section.
Move the general all-language guidance (General Principles, Return Values, Preprocessor Scripts) to a shared reference file (e.g. references/script-writing.md) referenced by each per-language skill, keeping only the CLI workflow and PostgreSQL section in SKILL.md.
Normalize heading levels (one H1, H2 sections) and add a short ordered checklist of the full flow: write → wmill generate-metadata → wmill script preview → deploy only on explicit request.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Nearly all content is Windmill-specific operational knowledge Claude does not already know (wmill commands, lock-hash mechanics, S3 directives), so there is no concept-explanation padding — but the same policies are repeated verbatim several times: "Only suggest/run a deploy when the user explicitly asks" (lines 15, 25-27, 45), and generate-metadata-is-not-a-deploy (lines 14, 33, 45). Sentences like the resource-type bullet and the metadata paragraph are long multi-clause run-ons that could be tightened. Mostly efficient with noticeable duplication to trim, matching anchor 3. | 3 / 5 |
Actionability | Fully executable throughout: exact commands (`wmill script preview <script_path>`, `wmill generate-metadata --dry-run`, `wmill generate-metadata rehash`, `wmill script preview <path> -d '<args>'`, `wmill resource-type list --schema`) and copy-paste-ready SQL (`SELECT * FROM users WHERE name = $1::TEXT AND age > $2::INT;`, `jsonb_to_recordset($1::jsonb) AS r(id INT, name TEXT)`, `-- s3 prefix=exports/users format=parquet`) cover the common cases, including per-language placeholder syntax. | 5 / 5 |
Workflow Clarity | Explicit decision rules sequence the workflow ("If the user says 'run the script'... use script preview"; "Only use sync push when...") and validation checkpoints are present — preview validates before any deploy, `--dry-run` lists stale items before narrowing, and the regenerated .lock files are diffed and version changes reported before deploying. Not 5 because the end-to-end sequence (write → generate-metadata → preview → deploy) is never laid out as one ordered flow; the reader must assemble it from four separate subsections, and the ordering of metadata regeneration vs. preview is only implicit. | 4 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are absent), so everything is inlined in one 114-line file. Structure exists (##/### sections) but heading levels are inconsistent ("# Windmill Script Writing Guide" and "# PostgreSQL" appear as H1s mid-document after top-level H2s), and general all-language Windmill guidance (General Principles, Return Values, Preprocessor Scripts — the latter explicitly TS/Python-only) is inlined in a PostgreSQL-specific skill where it belongs in a shared reference for the per-language skill family. Some structure, but content that should be separate is inline — anchor 3. | 3 / 5 |
Total | 15 / 20 Passed |