Conciseness

The body is dense with non-obvious, hard-won specifics (set-name conventions, tool parameters, failure modes) and contains almost no conceptual padding Claude already knows, so it is above anchor 2. But it could be tightened: the single-strand ORF caution is stated at length three times, 'final-answer discipline' is repeated across two sections, the six-traps summary re-states the detail sections below it, the gnomAD release table effectively appears twice (trap 6 and Limitations), and release-pinned values (gnomAD r4/r2.1 pLI, miRDB v6.0) are inlined rather than isolated — matching anchor 3's 'mostly efficient but could be tightened' rather than anchor 4.

Actionability

Fully executable throughout: exact tool names with parameters (`DisGeNET_get_disease_genes(disease=C0152200)`, `PubTator3_GetEntityRelations(e1="@GENE_<sym>")`), exact set-name patterns (`MIR186_3P`, `PGM3_TARGET_GENES`, `MP_INCREASED_MELANOMA_INCIDENCE`), and three complete, copy-paste-ready Python functions (count_orfs, digest, gamete_ratio) with worked examples and expected outputs. This is a direct match for anchor 5's 'copy-paste ready code or commands; specific examples cover the common cases'.

Workflow Clarity

Every multi-step process is a numbered sequence with explicit validation checkpoints and recovery loops: the 5-step MCQ procedure ('Check each option ... Exactly one option should be supported'), the 6-step in-X-but-not-Y recipe with a text-mined fallback, the 3-step drug-vignette procedure, and computational checks like 'Always check the returned length against what the question asked for', 'match the computed value back to the option letters carefully and explicitly', and 'If a tool errors on a name ... re-derive the name before concluding insufficient'. Feedback loops and checklists are present, matching anchor 5; no destructive/batch cap applies.

Progressive Disclosure

No bundle files exist (no references/, scripts/, or assets/), so the whole ~29KB — routing table, four domain playbooks (Allen, HPA, drug vignettes, gene-disease), and three code-heavy computational procedures — is inlined in one file. Internal structure is good (clear ## / ### headers, a routing table acting as an index, 'the detail for each is further down' signaling), which lifts it above anchor 2's 'minimal structure'; but content that clearly belongs in separate files (the executable Python in scripts/, the per-database playbooks in references/) is all inline, which is exactly anchor 3's 'content that should be separate is inline'. It cannot score higher absent any one-level-deep reference split.

Specificity

The description names concrete actions — "Answer biomedical FACTUAL / recall / multiple-choice questions by querying ToolUniverse database tools instead of answering from memory" and verifying "one option ... against an authoritative source" — plus named databases (DisGeNet, OMIM, MSigDB, ClinVar, etc.). It sits between anchor 3 and 5: the core capability is really one workflow (query tool → verify option → answer) rather than multiple distinct actions, so it does not reach anchor 5's 'multiple specific concrete actions', but it is noticeably above the 1-2-action midpoint.

Completeness

Both halves are explicit: the 'what' is the opening sentence (query ToolUniverse tools instead of answering from memory), and the 'when' is stated twice with concrete triggers — "Triggers on any 'which gene...' lookup" and "Use whenever a single correct answer exists in a public biomedical database and could be looked up rather than guessed". It also adds explicit NOT-for boundaries, matching the anchor-5 example's structure exactly. Anti-drift check: anchor 4's 'when could be more explicit' does not apply — triggers are concrete and repeated.

Trigger Term Quality

Trigger phrases are comprehensive and match what users would naturally say: "any 'which gene/drug/variant/disease/pathway/miRNA/TF...' lookup", "any question phrased 'according to <database>'", and 13 concrete database names plus "multiple-choice biology/medicine knowledge questions". Synonym coverage (factual / recall / multiple-choice / lookup) is complete; no natural phrasing of this task type is missing.

Distinctiveness Conflict Risk

The niche is unambiguous (biomedical database lookups via ToolUniverse) and it pre-empts the two adjacent skills by name: "NOT for analyzing user-supplied data files (CSV/VCF/h5ad → use the data-analysis router) and NOT for open-ended literature synthesis". Minimal overlap risk with any other skill.