Content
56%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 delivers genuinely actionable, executable guidance — real commands, flags, a routing table, and payload examples — and the workflow is clearly sequenced. However, a copy-paste artifact left the goldbox and best sections duplicated with broken numbering (1, 3, 4, 5, 2, 3, 4), and the trigger examples and purpose sections restate the description, wasting tokens and muddying navigation.
Suggestions
Deduplicate the Commands section: goldbox and best each appear twice (once as 'best --category', again as 'best-products --category-id') and the numbering runs 1, 3, 4, 5, 2, 3, 4 — collapse to a single numbered list of the five distinct subcommands.
Show how the 'vertical': 'book' and persona payloads are actually passed to the recommend command (a flag, a JSON file, or an inline argument), since the CLI examples never connect to those payloads.
Move the long trigger-phrase and category catalogs (which duplicate the frontmatter description) into a references/ file, keeping SKILL.md as a lean overview with the routing table and command examples.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Noticeable redundancy: the goldbox and best command sections each appear twice with full code blocks, section numbering runs 1, 3, 4, 5, 2, 3, 4, and both the "What this skill is for" bullets and the 14-item "Trigger examples" list largely restate the frontmatter description. This matches level 2 (several unnecessary or padded sections) rather than level 3, where only isolated tightening would be needed; roughly a quarter of the body is duplicative. | 2 / 5 |
Actionability | Commands are concrete and executable with real flags (--rocket-only, --max-price, --sort SALE, category lists, routing table), and the persona JSON payloads are copy-paste ready. It falls short of level 5 because the "vertical"/"persona" payloads are never connected to an actual CLI invocation — the recommend example shows only --query/--price-max/--limit, leaving unclear how the assist payload is passed. | 4 / 5 |
Workflow Clarity | The routing table plus 8 numbered steps give a clear intent-first sequence with sensible checkpoints ("ask for constraints if missing", prefer short_deeplink). These are read-only API calls so the destructive-operation validation cap does not apply, but explicit error recovery (what to do on {"ok": false}) is absent, matching level 4 (minor validation gaps) rather than level 5 (explicit feedback loops). | 4 / 5 |
Progressive Disclosure | Section headers and the routing table provide structure and the single referenced script exists, but the body inlines ~160 lines including duplicated command sections, two full trigger-phrase catalogs, and detailed persona/category reference material that belongs in a separate file. The duplicated and mis-numbered sections exceed the "minor organization gaps" of level 4, matching level 3 (content that should be separate is inline, organization could be better). | 3 / 5 |
Total | 13 / 20 Passed |