Content
63%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 retrieval workflow (Phases 0-3) is well sequenced, actionable, and backed by real tool calls, fallback chains, and error handling. However, the body bloats to ~300 lines by repeating the curation hierarchy multiple times and inlining six extended tool-catalog sections that belong in reference files, hurting both conciseness and progressive disclosure.
Suggestions
Move the extended tool catalogs (UniProt proteomes, UniParc/UniRef, BioSamples, BioThings, Bioregistry/Identifiers.org/TogoID, dbfetch) into reference files under references/, keeping one-line 'See X for Y' pointers in SKILL.md — this would cut the body by roughly half and fix progressive disclosure.
State the curation-quality hierarchy (RefSeq > RefSeq predicted > GenBank) exactly once — the Curation Level Tiers table — and delete the restatements in 'Domain Reasoning' and 'Reasoning Framework'.
Compress the 'verified live' narrative walkthroughs to the gotcha plus one compact example call/response pair, dropping the blow-by-blow test transcripts (e.g. the full SARS-CoV-2 proteome chain and EGFR UniParc narrative).
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient operational content, but with real padding: the curation-quality hierarchy (RefSeq > predicted > GenBank) is repeated four times (frontmatter, 'Domain Reasoning', 'Reasoning Framework', 'Curation Level Tiers'), and the long 'verified live' narrative transcripts (e.g. the SARS-CoV-2 proteome chain and EGFR UniParc walkthroughs) recount full test runs where the gotcha plus one compact example would suffice. This fits 'Mostly efficient but includes some unnecessary explanation or could be tightened'. | 3 / 5 |
Actionability | Concrete tool calls with real parameter names and formats (tu.tools.NCBI_search_nucleotide with organism/gene/strain/keywords/seq_type/limit), explicit fallback chains, an error-to-response table, and real example invocations with actual outputs. Minor gaps keep it below fully copy-paste-ready: snippets use unbound placeholders (accession=accessions) with no complete end-to-end runnable example. | 4 / 5 |
Workflow Clarity | Clear sequenced phases (Clarify → Disambiguate → Retrieve → Report) with an identity checklist, fallback chains, and an error-handling table forming feedback loops. Falls short of the top anchor because verification of the retrieved sequence (e.g. confirming organism/strain matches the request, per Synthesis Question 4) is posed as a question rather than an explicit checkpoint inside the workflow. | 4 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are all absent), so everything lives in a ~300-line SKILL.md. Section headers are clear, but roughly half the body — the extended per-tool catalogs (UniProt proteomes, UniParc/UniRef, BioSamples, BioThings, Bioregistry/TogoID, dbfetch) — is reference material that clearly belongs in separate files with one-line pointers. This matches 'Some structure but content that should be separate is inline'. | 3 / 5 |
Total | 14 / 20 Passed |