Conciseness

The core content (gotchas on strand-aware cis-windows, ALT-effect beta, duplicate-rsid collapsing) is genuinely non-obvious domain knowledge, but several passages are heavily padded: a full inline citation for Sun 2018 (doi, PMID, author list) in the Trigger section, cache-probing statistics ('33 of 758 probed 2026-09-13', 'measured 2026-10-22, 0 of 468 repeated keys disagreed (82 in the SORT1 QTD000276 window, 311 in QTD000609, 75 in QTD000269 .cc)'), and time-sensitive dates (HTTP 410 'since September 2026') not placed in a deprecated section, which the rubric explicitly penalizes. Gotchas 6 and 8 are 200+ word run-on paragraphs that could be tightened substantially.

Actionability

The CLI Reference gives copy-paste-ready commands covering standard usage, the bundled demo, demo listing, and the ClawBio runner variant, plus a concrete JSON config schema and realistic example output (manifest.yaml, variants.tsv, collapsed_rows.tsv, report.md). Fully executable guidance with specific examples covering the common cases.

Workflow Clarity

The 5-step workflow (resolve dataset_id -> pick region -> tabix range fetch -> filter by molecular_trait_id -> write outputs) is clearly sequenced with concrete commands, and validation is explicit and thorough: every row is checked against the 19 expected columns with EQTLCatalogueSchemaError naming the file and bad row, the CLI exits 2 and writes nothing on failure, duplicate-stat conflicts raise EQTLCatalogueDuplicateRowConflict, and the invariant 'n_variants + len(collapsed) == n_rows_read is checked, not assumed'. No validation cap applies.

Progressive Disclosure

Sections are well-organized with clear headers, but everything lives in one ~186-line monolithic SKILL.md: the enormous Gotcha 6 and 8 paragraphs, the full example-output dumps, and the citation detail are content that clearly belongs in separate reference files. Referenced bundle paths (data/dataset_index_r7.tsv, examples/input.json, eqtl_catalogue_region_fetch.py) are not present in the bundle to verify, and no dedicated reference files carry any of the detail.

Specificity

'Fetch a region of cis-eQTL summary statistics from EBI eQTL Catalogue v7+ via tabix-on-FTP' plus enumerated inputs (dataset_id, chromosome, start, end, optional molecular_trait_id) and a concrete output ('harmonised TSV slice') gives multiple specific concrete actions with comprehensive coverage. Not 4: there are no gaps — the mechanism, the per-variant fields (beta / SE / p-value), and the input/output contract are all stated.

Completeness

Both 'what' ('Fetch a region of cis-eQTL summary statistics... via tabix-on-FTP') and 'when' ('Use when an agent needs eQTL beta / SE / p-value for every variant in a window around a gene's TSS for one specific dataset') are explicitly and concretely answered, with the input/output contract on top. Not 4: the 'when' clause is fully explicit, not merely present.

Trigger Term Quality

Strong domain-natural terms ('cis-eQTL', 'eQTL Catalogue', 'eQTL beta / SE / p-value', 'window around a gene's TSS', 'study × tissue × quantification method') that a bioinformatician would naturally say, but common variations like 'eQTLs for a gene', 'summary stats', 'colocalisation', or cohort names like 'GTEx' are missing. Not 3: keyword coverage goes well beyond 'some relevant keywords'; not 5: a few natural synonyms are absent.

Distinctiveness Conflict Risk

The description carves a clear niche (one dataset, one region, tabix-on-FTP from EBI eQTL Catalogue v7+) that would not fire for sibling skills such as single-variant GTEx Portal lookups, UKB-PPP pQTL fetches, or fine-mapping posteriors. Not 4: overlap risk is minimal because the scope is narrowed to one (study × tissue × quantification method) regional fetch.