PubMed search finds papers. It does not read them, sort them, or tell you whether they agree. This workflow does that part, using Claude’s PubMed connector.
I did not build the connector. It is an off-the-shelf integration that gives the model direct access to PubMed. What I built is the workflow on top of it: how to scope a question, build the query, screen results, extract an evidence table, and write a synthesis that flags where studies conflict.
What it does
- Search. Builds a PubMed query from a plain-language question, with synonyms and field tags, and pulls every result for the date range.
- Screen. Reads every title and abstract and sorts them into relevant, adjacent, and off-topic, so you only read what matters.
- Expand. Follows related-article links from the key papers to catch work the query missed.
- Tabulate. One row per relevant paper: study type, population, exposure, outcome, key finding, limitation, PMID, and DOI.
- Synthesize. A short memo grouped by question, with the agreements, the conflicts, and what they mean for exposure and risk assessment.
- Alert. A scheduled rerun that screens only new papers and sends a digest when something relevant appears.
A test run
On 1,4-dioxane, 2023 to 2026, the search returned 73 papers. Of the first 40, a quarter were off-topic, mostly studies that use dioxane as a lab solvent. The synthesis surfaced a live disagreement over dioxane’s cancer mode of action that matters for how it is regulated.
Limits
PubMed only covers biomedical and life sciences literature, screening is based on abstracts unless full text is pulled, and every finding needs a human check before it goes into a report.