Content
75%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-structured, highly actionable instruction-only router skill with executable commands, clear sequencing, guardrails against fabrication, and a properly signaled one-level-deep reference. Its main weaknesses are redundancy — catalog sources listed twice and taxonomy content duplicated between the body and the reference — plus an unspecified step for checking whether a skill is already installed.
Suggestions
Deduplicate the catalog source list: state the four sources once and reference that list from the fallback instruction instead of repeating all four URLs verbatim.
Move the catalog-lane enumeration and distinctive-phrase lists out of the body into references/taxonomy-routing.md (which already contains them), keeping only a pointer plus a few examples in SKILL.md.
Specify how to perform step 1 (e.g., check available skills or in-context skill names for an installed NVIDIA skill before querying the catalog).
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The four catalog sources are listed twice verbatim (the 'Catalog check means one of' bullets and again as 'Use the fallback catalog sources' bullets), and the product/lane enumerations repeat the description and duplicate references/taxonomy-routing.md. It is mostly purposeful with no explanations of concepts Claude already knows, matching anchor 3 ('could be tightened') rather than 2, but the duplicated passages keep it below 4. | 3 / 5 |
Actionability | Fully executable, copy-paste-ready guidance covering the common cases: 'npx skills add nvidia/skills --list' with three concrete fallback URLs, per-agent install commands ('--agent codex', '--agent claude-code', and omitting --agent when unknown), a literal recommendation-format template, and a defined failure path ('say the catalog could not be checked and do not name a concrete slug'). Matches anchor 5. | 5 / 5 |
Workflow Clarity | A clear numbered 5-step discovery sequence (check installed → query live catalog → match → recommend at most three → ask before installing) with error-recovery paths: fallback sources when the CLI fails and anti-fabrication handling when lookup fails. Minor validation gaps — how to check whether a skill is 'already installed or already in context' is unspecified and there is no post-install verification — place it at anchor 4 rather than 5. | 4 / 5 |
Progressive Disclosure | One reference file (references/taxonomy-routing.md, which exists in the bundle) is clearly signaled with a conditional load instruction ('Read ... only when the request is taxonomy-only, ambiguous') and is one level deep. However, the body inlines roughly 25 lines of catalog-lane and distinctive-phrase lists that duplicate the reference's 'Stable Catalog Lanes' content, so content is not fully split — anchor 4 rather than 5. | 4 / 5 |
Total | 16 / 20 Passed |