Content
52%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 core Phases 0–3 workflow is well-sequenced and actionable with concrete tool calls, fallbacks, and error handling, but the body is bloated: the curation hierarchy is repeated four times and half the file is a tangential tool catalog with verbose "verified live" anecdotes that should live in reference files. Splitting the tool catalog into references/ and de-duplicating the hierarchy would dramatically cut token cost without losing clarity.
Suggestions
Move the UniProt/UniParc/UniRef/BioSamples/BioThings/Bioregistry/TogoID/dbfetch sections into references/*.md files and keep only a one-line "Use when..." pointer per tool in SKILL.md, fixing progressive_disclosure and cutting the body roughly in half.
State the RefSeq (NM_/NP_) > predicted (XM_/XP_) > GenBank curation hierarchy exactly once (e.g., in the "Domain Reasoning" or "Curation Level Tiers" section) and delete the redundant restatements in "Reasoning Framework" and "Accession guidance".
Compress the "verified live" narrative examples into short factual bullets (input → key output → gotcha), keeping the genuinely non-obvious quirks (e.g., "Cell membrane" vs "plasma membrane", ENA 404 on RefSeq) while trimming the step-by-step recounting.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is noticeably verbose in several places: the RefSeq>predicted>GenBank curation hierarchy is stated four times (frontmatter-echo in "Domain Reasoning", "Reasoning Framework", "Accession guidance", and the "Curation Level Tiers" table), and lines ~154–305 are a ~150-line catalog of tangential tools (UniParc, UniRef, BioSamples, BioThings, Bioregistry, TogoID, dbfetch) padded with long "verified live" narrative anecdotes. This fits the 2 anchor (several unnecessary/padded sections) rather than 3's "some unnecessary explanation or could be tightened". | 2 / 5 |
Actionability | Concrete, mostly executable guidance: real tool calls with parameters ("tu.tools.NCBI_search_nucleotide(operation="search", organism=..., limit=10)"), a parameter reference ("format (fasta/genbank)"), fallback chains, and an error→response table with verified real outputs. Not 5 because the tooluniverse/"tu.tools" context is never established and result-structure assumptions like result["data"]["uids"] cannot be verified as copy-paste ready. | 4 / 5 |
Workflow Clarity | A clear Phase 0→3 sequence with explicit skip conditions, an identity checklist, fallback chains, and an error-handling table, plus the "LOOK UP DON'T GUESS: Never assume accession numbers... Always retrieve and verify" directive. It fits 4 (clear sequence, most checkpoints present) rather than 5 because recovery is table-based rather than an explicit validate→fix→retry loop, and rather than 3 because checkpoints and fallbacks are explicitly present. | 4 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, or assets/), and the entire ~305-line body is inline — including a "Tool Reference", "Search Parameters Reference", and ~150 lines of tangential Uni*/dbfetch tool documentation that clearly belongs in separate reference files, with zero pointers to any external file. This matches the 2 anchor (content that clearly belongs in separate files is inlined) rather than 3, since there are no separate-file references at all, only inline bulk. | 2 / 5 |
Total | 12 / 20 Passed |