pubmed

Files (20)

• SKILL.md
• openai.yaml · agents/
• caller-contract.md · references/
• cli.md · references/
• eutils.md · references/
• mcp.md · references/
• pubmed.py · scripts/
• __init__.py · scripts/pubmed_tool/
• __main__.py · scripts/pubmed_tool/
• cli.py · scripts/pubmed_tool/
• client.py · scripts/pubmed_tool/
• commands.py · scripts/pubmed_tool/
• errors.py · scripts/pubmed_tool/
• storage.py · scripts/pubmed_tool/
• xml_parser.py · scripts/pubmed_tool/
• citations.json · tests/fixtures/
• esearch-warning.json · tests/fixtures/
• pubmed-records.xml · tests/fixtures/
• test_client.py · tests/
• test_pubmed_tool.py · tests/

SKILL.md (first lines)

# PubMed

Use PubMed as a reproducible evidence lookup, not a title generator. Keep scientific judgment in the agent and route retrieval by required capability.

## Define the search

Clarify the information need and whether the search is a rapid lookup, scoping search, citation check, or systematic-search contribution. Translate clinical questions into PICO/PECO/PICOS concepts where useful.

For each concept, combine:

- controlled vocabulary, including current MeSH terms;
- title/abstract synonyms, spelling variants, acronyms, and older terminology;
- study-design terms only when they serve the stated purpose.

Build Boolean blocks explicitly. Inspect PubMed's query translation or Search Details when available. Treat field tags as exact constraints because they turn off Automatic Term Mapping for the tagged term.

The search definition is complete when another researcher could rerun the exact query and understand every applied limit.

## Retrieve and verify

When another skill invokes PubMed, use the stable [caller contract](references/caller-contract.md). Keep backend selection, response accounting, and artifact handling inside this skill so callers remain portable across Claude Code and other agent hosts.

If PubMed MCP tools are available, route the request according to [references/mcp.md](references/mcp.md). Use the bundled Python CLI according to [references/cli.md](references/cli.md) for evidence-grade retrieval and verification. Use raw NCBI E-utilities according to [references/eutils.md](references/eutils.md) only as the final fallback.

Give either search path an already-formed query and inspect its translation before accepting the search. The CLI exposes a zero-network `capabilities` manifest plus `search` for evidence-grade retrieval, `fetch` for authoritative records, `verify` for deterministic bibliographic checks, `related` for PubMed neighbors, `cite-match` for structured citation resolution, and `id-convert` for PMID/PMCID/DOI conversion. The MCP adds bounded lookup, PMC full text, and copyright/licence status. Neither path chooses terms, screens studies, appraises evidence, or decides whether a citation supports a claim.

Retrieve full PubMed records for candidate PMIDs. Verify title, authors, journal, publication year, abstract, publication type, MeSH terms, DOI, correction/retraction links, and PMID from the record itself. Deduplicate by PMID first, then DOI and normalized title.

For large searches, fetches, verifications, or related-record expansions, request durable artifacts with summary-only stdout. Read targeted records from the returned artifact paths instead of placing the complete payload in the agent context.

For topic refinement, first locate recent systematic reviews and protocols, then use their terminology and included studies to find landmark and newer primary reports. For claim verification, read enough of the abstract or accessible full text to determine whether the cited paper actually supports the claim.

Retrieval is complete only when every requested PMID is accounted for, every reported paper has a verified PMID, and the result set's coverage limits are stated. A result over 10,000 has no complete ESearch PMID manifest until explicit non-overlapping partitions are supplied and validated.

## Report reproducibly

Return:

1. the exact PubMed query;
2. search date, result count, sort order, and applied limits;
3. a concise synthesis answering the research need;
4. a table or list with title, first author, year, journal, study type, PMID, DOI when present, and `https://pubmed.ncbi.nlm.nih.gov/{PMID}/`;
5. caveats about indexing, unavailable abstracts/full text, retractions or corrections, and likely missed terminology.

Label inference as inference. A PubMed-only search is not an exhaustive systematic search; recommend additional databases and information sources when the user is conducting a systematic review or meta-analysis.

From research-consultancy/rc-skills, MIT.

More from research-consultancy/rc-skills