How to Use the BioSkepsis Connector in Claude for Biomedical Research
Reviewed
How to Use the BioSkepsis Connector in Claude for Biomedical Evidence Questions
Connecting BioSkepsis to Claude takes four clicks; using the BioSkepsis connector well takes a little more. This post covers how to phrase a biomedical question so a BioSkepsis research run accepts it, which of the five output formats to request, how to check the verification signals on the brief that comes back, and what to do with a brief once Claude delivers it. Every step is shown on one real run about anti-amyloid antibodies in early Alzheimer's disease, and the full thread from that run is linked below so every figure in this post can be checked at source.
Why Send a Biomedical Evidence Question Through the BioSkepsis Connector Instead of Claude Alone
Claude answers a biomedical question from training data or from a general web search when no literature connector is switched on in the Claude conversation. The BioSkepsis connector changes where the citations in a biomedical answer come from, and changes what happens to each citation before a researcher reads it.
A BioSkepsis research run searches a curated corpus of more than 40 million biomedical papers published from 1931 onward, reads the full text of the retrieved papers where full text is available, and checks every claim in the draft brief against the paper that the claim is attributed to. Claims that fail the citation check are struck from the BioSkepsis brief and listed separately, so a researcher can see what was removed rather than trusting that nothing was removed.
The cost of that verification work is time. A BioSkepsis research run takes minutes, while Claude alone answers a biomedical question in seconds. For a quick definition of a term, Claude alone is the faster choice; for a claim that will enter a grant application, a manuscript, a clinical protocol or a tumour-board note, a verified BioSkepsis brief is the safer input.
| Dimension | Claude with the BioSkepsis connector | Claude with no literature connector |
|---|---|---|
| Where the evidence comes from | More than 40 million curated biomedical papers, published from 1931 onward | Claude training data, or general web search results |
| How deeply each paper is read | Full text where full text is available, including methods and supplementary data | Whatever text a web result or training data contains |
| Handling of retracted and corrected papers | Retracted and questioned papers are screened, and corrections are reported in the provenance facet | Retractions and corrections are not systematically checked |
| Checking of each individual citation | Every cited claim is checked against its source paper, and failed citations are struck and listed | No claim-by-claim citation check runs |
| Rating of the evidence base | A Trust Index with grounding, provenance, evidence strength, reproducibility and durability facets | No evidence rating is produced |
| Reporting of what was not answered | A research notebook reports each sub-question, the statements behind it, and a verification status | Gaps in the answer stay implicit |
| Time to an answer | Usually 3 to 8 minutes per research run, sometimes longer | Seconds |
| Cost of an answer | One monthly BioSkepsis research brief per research run | Included in the Claude plan |
How to Write a Biomedical Question That a BioSkepsis Research Run Will Accept
The BioSkepsis connector is scoped to life-science and biomedical research, and a question outside that scope returns a declined status with a short explanation instead of a brief. Coding questions, finance questions and one-line factual lookups are declined, because a full literature search would add nothing to a question that needs no literature.
Inside that scope, the shape of the question decides the quality of the brief. A question that names the population, the intervention or mechanism, and the outcome gives the BioSkepsis run enough structure to split the question into sub-questions and to report a verification status for each sub-question separately. A bare keyword string gives the BioSkepsis run no outcome to search against and no population to filter on.
A question the BioSkepsis connector cannot work with
The prompt "anti-amyloid antibodies Alzheimer" names a drug class and a disease, but the prompt names no population, no comparator and no outcome, so a BioSkepsis research run has nothing specific to verify claims against.
The same topic, written as a full biomedical research question
The question "Do anti-amyloid monoclonal antibodies such as lecanemab and donanemab produce clinically meaningful cognitive benefit in early Alzheimer's disease, and how do the reported benefits compare with the rates of amyloid-related imaging abnormalities?" names two drugs, a disease stage, a benefit outcome and a harm outcome, which is why the BioSkepsis run split the question into six sub-questions covering the size of the benefit for each drug, the rates and consequences of brain-swelling events, the factors that raise those rates, the clinical meaningfulness judgement, and the independent critiques of the benefit-to-harm balance.
Asking a two-part question, as the anti-amyloid example does, is usually better than asking two separate questions, because each BioSkepsis research run uses one monthly research brief. A benefit question and a harm question answered inside one run also share the same retrieved papers, which makes the benefit numbers and the harm numbers directly comparable.
Choosing an Output Format for a Biomedical Literature Brief in Claude
The BioSkepsis connector accepts an optional output format on every research run, and the five output formats are answer, review, interpretation, summary and extraction. The answer format is the default and returns a direct answer to the question that was asked, the review format returns a literature review, the interpretation format interprets a finding, the summary format returns a summary, and the extraction format returns detail extracted from the retrieved papers.
Requesting an output format in Claude needs no special syntax: a phrase such as "run this as a literature review in BioSkepsis" is enough for Claude to pass the review format to the BioSkepsis connector. Leaving the output format out is a reasonable default, because the answer format suits most questions that ask whether something works, how large an effect is, or how two options compare.
Choosing an output format is a choice about the shape of the document, not about how hard the BioSkepsis run works. Citation verification, retraction screening, sub-question reporting and the Trust Index run on every BioSkepsis research run regardless of which output format was requested.
A Worked Run: Lecanemab and Donanemab in Early Alzheimer's Disease
The anti-amyloid question above was run through BioSkepsis on 15 September 2026, and the full thread is public at the research link above. The BioSkepsis run answered the benefit question with a direct no: the benefit of lecanemab and of donanemab is statistically significant but small in absolute terms, and falls below the published thresholds for a minimal clinically important difference, while brain-swelling events rise severalfold above placebo.
The benefit figures behind that answer are specific. Lecanemab worsened by 1.21 points against 1.66 points on placebo over 18 months on the Clinical Dementia Rating Sum of Boxes scale, a clinician-rated scale of cognitive and functional impairment, which is 0.45 points less decline and was reported as 27% slowing (PMID 38951040). Donanemab differed from placebo by 2.92 points in the combined population and 3.25 points in the low-tau and medium-tau population at 76 weeks on the integrated Alzheimer's Disease Rating Scale, with Clinical Dementia Rating Sum of Boxes differences of 0.7 and 0.67 points, and 23 of 24 pre-specified outcomes were statistically significant (PMID 37459141).
The judgement on meaningfulness comes from sources outside the trial reports, which is the part a general-purpose assistant is least likely to supply. Published thresholds for a minimal clinically important difference are roughly 1.0 to 1.6 points on the Clinical Dementia Rating Sum of Boxes scale and 5 to 9 points on the integrated Alzheimer's Disease Rating Scale, so the observed group-level differences fall short of both thresholds (PMID 38951040). A 2026 Cochrane review pooling 17 phase 3 trials graded the 18-month class effects as trivial, at roughly 0.29 points on the Clinical Dementia Rating Sum of Boxes scale and 0.85 points on the Alzheimer's Disease Assessment Scale cognitive subscale, and concluded that the drug class does not offer a favourable balance of benefit against risk in mild cognitive impairment or mild dementia (PMID 41985900).
The harm side of the same run, reported in matched units
Brain swelling detected on imaging, the main safety concern with anti-amyloid antibodies, occurred in 12.6% of lecanemab-treated participants against 1.7% on placebo, and was symptomatic in 2.8% (PMID 38730496), while donanemab produced brain swelling in 24.4% against 1.9% on placebo, symptomatic in 5.8%, with serious events in 1.5% (PMID 40063015). Pooled across the drug class, the 2026 Cochrane review put the risk ratio for any brain swelling at 10.02 against placebo, which is about 107 extra events per 1,000 people treated, set against a mean benefit of roughly 0.29 points on the Clinical Dementia Rating Sum of Boxes scale (PMID 41985900).
A regulatory split the same run surfaced without being asked for it
Regulators and payers read the same trial data differently: the United States Food and Drug Administration granted traditional approval to lecanemab and approved donanemab in 2024, and the European Medicines Agency and the United Kingdom Medicines and Healthcare products Regulatory Agency licensed lecanemab, while the United Kingdom National Institute for Health and Care Excellence rejected reimbursement of both drugs, the Dutch Zorginstituut refused reimbursement of lecanemab for absent net benefit, and the German Institute for Quality and Efficiency in Health Care concluded that lecanemab offers no added benefit (PMID 41985900).
A bias risk the run quantified rather than merely mentioning
Because brain swelling and infusion reactions are far more common in the treated arms, participants and carers may guess which arm they are in, which can bias carer-rated scales, and the 2026 Cochrane review downgraded the certainty of the efficacy results for that reason (PMID 41985900). A statistical analysis of the same problem estimated that fully explaining the observed effects through unblinding alone would need therapeutic-insight effects of 3.7 points for lecanemab and 3.3 points for donanemab, which is larger than plausible, so unblinding is unlikely to account for the whole effect while still possibly explaining part of it (PMID 38380503).
A researcher who asked the same anti-amyloid question of a general-purpose assistant would plausibly get the two headline trial results and a sentence saying the benefit is debated. The BioSkepsis run instead returned the benefit figures, the thresholds those figures are measured against, the harm rates in matched units, the genetic and imaging risk factors, the monitoring schedule from the donanemab appropriate-use recommendations (PMID 40155270), the regulatory split, a quantified estimate of unblinding bias, and a list of what the literature still does not establish.
How to Check the Verification Signals on a Biomedical Brief Before Quoting Any Number
A BioSkepsis brief is built to be audited rather than trusted on sight, and the audit trail sits in the research notebook at the end of the thread. The notebook lists each sub-question, then lists the individual statements that answer that sub-question, and each statement names the PubMed identifier and the section of the paper it was taken from, such as the results section or the discussion section.
Each sub-question in the anti-amyloid thread carries a verification status and a confidence level, and all six sub-questions in that thread are marked as verified with high confidence. A researcher who wants to quote the donanemab brain-swelling rate can therefore open the notebook entry for the harm sub-question, read the statement, and go straight to the cited paper rather than to the summary text at the top.
The Trust Index adds the second layer of checking, with separate facets rather than one score. The grounding facet counts how many cited supports were verified against full text and how many were struck after failing verification, the provenance facet flags retractions, corrections and venue issues on the cited papers, the evidence strength facet reflects the study designs behind the findings, the reproducibility facet reflects how many independent research groups support each load-bearing finding, and the durability facet reflects how recently the evidence base has been active.
What the anti-amyloid run flagged about its own sources
The anti-amyloid brief notes that both pivotal trials were funded by the companies selling the drugs, that the relative-slowing percentages originate from those sponsor analyses, and that two papers from one sponsor about that sponsor's product are a single source of evidence rather than independent corroboration (PMID 41985900). The same brief notes that a sponsor-funded safety analysis reporting no adverse effect of brain swelling on cognition should be weighted as a sponsor analysis of its own product (PMID 38730496).
What to Do With a Biomedical Brief After Claude Delivers It
A finished BioSkepsis brief inside a Claude conversation is a starting point for the rest of the work, not an end point. Claude can quote the brief into a draft section, pull the PubMed identifiers into a reference list, or compare the brief against a draft document already open in the Claude conversation.
Two of the six BioSkepsis tools exist for reuse rather than for new research. The list_recent_briefs tool lists recent BioSkepsis briefs, and the get_brief tool re-opens a saved BioSkepsis brief by its identifier, and neither tool starts a new research run or uses another monthly research brief.
The create_research_feed tool turns a finished research run into a standing alert: asking Claude to follow the topic calls the create_research_feed tool with the run identifier and a daily, weekly or monthly frequency, and BioSkepsis then emails newly published papers related to the papers the run used. A standing alert is useful on this particular topic because the real-world safety picture is still forming: a Japanese all-case surveillance reported brain swelling in 7.1% of lecanemab-treated patients over about six months (PMID 41849958), while a United States specialty clinic reported 22% over 6.5 months (PMID 40354064), and both reports cover short follow-up periods.
Who Gains Most From the BioSkepsis Connector in Clinical and Biomedical Work
ConnectorClinicians and medical writers checking a drug claim
A clinician or medical writer who needs one defensible answer about a drug should ask one full biomedical question in a Claude conversation, then open the research notebook and read the cited statement before pasting any PubMed identifier into a document. The anti-amyloid run shows why that check matters, because the same trial numbers support a 27% slowing headline and a judgement that the absolute difference sits below the threshold for a minimal clinically important difference.
ConnectorPhD students and postdocs drafting a literature review
A PhD student or postdoc drafting a review chapter should request the review output format and keep the research notebook, because the notebook states which sub-questions the retrieved literature actually answered and with what confidence. The evidence gaps section of a BioSkepsis brief also names what no study has yet established, which is the material a good review chapter needs for its final paragraphs.
ConnectorPharmaceutical and translational research teams
A pharmaceutical or translational research team can add the BioSkepsis connector across an organisation so that each member signs in with their own BioSkepsis account and keeps their own brief allowance. Research feeds created from finished runs then track new publications per target or per indication without anyone starting a new research run each month.
Frequently asked questions
What kind of biomedical question should be sent to the BioSkepsis connector in Claude?
The BioSkepsis connector works best on a full biomedical question that names a population, an intervention or a mechanism, and an outcome, for example whether lecanemab slows cognitive decline in early Alzheimer's disease. The BioSkepsis connector declines questions outside life-science and biomedical research, and declines quick factual lookups that need no literature search.
How long does a BioSkepsis research run started from Claude take?
A BioSkepsis research run usually takes 3 to 8 minutes and can take longer on a broad biomedical question, because the run searches the literature, reads full text where full text is available, and verifies each cited claim against the paper it is attributed to. Claude reports the phase of the run while the run is in progress and posts the brief when the run is finished.
What do the five BioSkepsis output formats do?
The BioSkepsis connector offers five output formats: answer, review, interpretation, summary and extraction. The answer format returns a direct answer and is the default, the review format returns a literature review, the interpretation format interprets a finding, the summary format returns a summary, and the extraction format returns extracted detail from the retrieved papers.
How can a researcher check that a BioSkepsis brief is grounded in the papers it cites?
A BioSkepsis brief carries a research notebook that lists each sub-question, the individual statements drawn from the retrieved papers, the PubMed identifier and paper section behind each statement, and a verification status for each sub-question. The BioSkepsis brief also carries a Trust Index whose facets cover grounding, provenance, evidence strength, reproducibility and durability, so a researcher can see how much of the brief was verified against full text before quoting any number.
Does a BioSkepsis research run started in Claude use a monthly research brief?
Every BioSkepsis research run started from Claude uses one of the monthly research briefs included in the BioSkepsis account that authorised the connector. Re-opening a saved BioSkepsis brief with the list_recent_briefs tool or the get_brief tool does not start a new run and does not use another monthly research brief.
Can Claude keep following the literature after a BioSkepsis brief is delivered?
Claude can create a BioSkepsis research feed from a finished research run by calling the create_research_feed tool with the run identifier and a daily, weekly or monthly frequency. The BioSkepsis research feed then emails newly published papers related to the papers that the finished research run used.
Run Your Next Biomedical Evidence Question Through BioSkepsis in Claude
Create a BioSkepsis account, add the BioSkepsis connector in Claude, and ask the biomedical question you were about to ask anyway. Every BioSkepsis brief arrives with PubMed citations checked against their source papers and a research notebook you can audit statement by statement.
Start freeSources & further reading
- PMID 38951040 — Critical reanalysis of the anti-amyloid antibody effect sizes, source of the lecanemab change of 1.21 against 1.66 points, the 0.45-point difference reported as 27% slowing, and the thresholds for a minimal clinically important difference.
- PMID 37459141 — TRAILBLAZER-ALZ 2 report of donanemab in early symptomatic Alzheimer's disease, source of the differences of 2.92 and 3.25 points on the integrated Alzheimer's Disease Rating Scale and of the 23 of 24 statistically significant outcomes.
- PMID 41985900 — 2026 Cochrane review of 17 phase 3 trials, source of the pooled effects of about 0.29 and 0.85 points, the risk ratio of 10.02 for brain swelling, the regulatory and reimbursement split, and the unfavourable benefit-to-risk conclusion.
- PMID 38730496 — Updated lecanemab phase 3 safety results, source of the lecanemab brain-swelling rates of 12.6% against 1.7% on placebo and 2.8% symptomatic.
- PMID 40063015 — Pooled analysis of amyloid-related imaging abnormalities with donanemab, source of the donanemab rates of 24.4% against 1.9% on placebo, 5.8% symptomatic and 1.5% serious.
- PMID 40155270 — Donanemab appropriate-use recommendations, source of the imaging monitoring schedule and of the prescribing cautions on anticoagulants and APOE4 status.
- PMID 38380503 — Analysis of the potential impact of unblinding on observed treatment effects, source of the estimated 3.7-point and 3.3-point therapeutic-insight effects.
- PMID 41849958 — Japanese all-case surveillance of lecanemab, source of the real-world brain-swelling rate of 7.1% over about six months.
- PMID 40354064 — United States specialty clinic cohort of lecanemab, source of the real-world brain-swelling rate of 22% over 6.5 months.
- BioSkepsis research thread — the full anti-amyloid run behind every figure in this post.
- BioSkepsis blog — Claude Connection for Biomedical Research: How to Connect Claude, the setup steps for the BioSkepsis connector.
- BioSkepsis documentation — BioSkepsis Claude connector documentation, the full reference including plan allowances.