More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    BGPT Paper Search: Claim-Level Evidence From Full Text

    1

    Find claim-level evidence in scientific literature through full-text paper search: craft answerable questions, read extractions skeptically, verify numbers, build evidence tables.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    BGPT Paper Search: Claim-Level Evidence From Full Text

    BGPT Paper Search: Claim-Level Evidence From Full Text

    Example session with this skill installed

    Find claim-level evidence for "Metformin lowers HbA1c in adults with type 2 diabetes." I need the effect size, the populations studied, and the limitations the authors declared. Synthesize it honestly, contradictions included.

    • Read your context and instructions
    • Compiled the bgpt paper search

    EVIDENCE SYNTHESIS — metformin and HbA1c: real effect, heterogeneous, with the contradictions kept in.

    Queries (variants, deduplicated by DOI): "metformin HbA1c randomized trial," "metformin glycemic control adults type 2 diabetes," "metformin HbA1c reduction effect size." Filters: primary studies and meta-analyses, last 10 years, 20 results per variant. 60 results → 34 unique papers by DOI. Query history recorded for reproducibility.

    Reading skeptically: two records are legacy/metadata records with empty evidence lists — excluded from synthesis, noted as such. One extraction reports an HbA1c reduction of 1.1%, but the source paper's table says 0.9% with a different unit convention — the extraction normalized incorrectly. The number is corrected to the source value and the correction noted in the provenance column. Three papers report contradictory results; all three stay in the table.

    Evidence table (abridged):

    • DOI 10.xxxx/a — RCT, 24 weeks, n=412: −0.9% (95% CI −1.1, −0.7). Scope: adults with type 2 diabetes, single country. Limitation (authors note): short follow-up. Provenance: Results, Table 2.
    • DOI 10.xxxx/b — RCT subgroup, n=188: −0.4% (95% CI −0.8, 0.0). Scope: adults 65+. Limitation (authors note): subgroup, underpowered. Provenance: Results, Fig. 3.
    • DOI 10.xxxx/c — RCT, n=96: −0.2% (95% CI −0.6, 0.2). Scope: adults, one clinic. Limitation (authors note): small n, single site. Provenance: Results, Table 1.

    Synthesis: the effect is real but heterogeneous — around −0.9% in general adult populations, smaller in elderly patients, and absent in one small single-site trial. The scope column is what prevents overgeneralizing the headline number to every patient. The contradictory trial is reported, not dropped — it is the reason the synthesis says "heterogeneous" instead of "proven."

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Abstracts summarize; full text evidences. A claim-level search index extracts what papers actually did — experiments, reported statistics, scope, limitations — so you can synthesize evidence instead of collecting citations. The skill is using it without outsourcing your judgment to the extraction.

    It starts with fit: the index is right for "what experiments support claim X," "what numbers were actually reported," and "did this finding replicate" — and wrong for systematic reviews, risk-of-bias assessment, or understanding a key paper (read the paper for that). Searching is questions, not keywords: short, specific English queries naming the phenomenon and the population or intervention, with date ranges, study types, and result counts set as explicit filters instead of stuffed into the query text. There is no pagination, so coverage comes from explicit question variants deduplicated by DOI, with the query history kept for reproducibility. Papers you already know get looked up by DOI directly.

    Reading results skeptically is the core: check the extraction status (a returned record doesn't guarantee claim-level provenance), respect the bounds (extractions are pointers into the paper, not the paper), verify every number against the source (normalization introduces errors), separate author-declared limitations from extraction commentary, and keep contradictory evidence in the table — dropping the inconvenient study is how reviews become marketing. The evidence table has fixed columns (DOI, claim, experiment/design, n, key statistic, scope, limitations, provenance), and the scope column is what stops you from overgeneralizing later. Budget hygiene covers metered pricing, deliberate result counts, and the standing rule: a connection error is not an empty evidence set.

    • It is a metal detector, not an excavation: extractions point into papers; they don't replace reading the key ones.
    • It cannot enumerate the literature exhaustively — it is not a systematic-review search.
    • Extraction normalizes formats, and normalization introduces errors; every number in the synthesis must be verified against the source.
    • A record with no evidence units is a metadata record, not a claim-level extraction, and must be excluded from synthesis.
    • Metered APIs charge per returned result; large trawls cost real money and need a deliberate budget.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 3 days ago

    Permissions

    No special permissions declared or detected

    Listed3 days ago

    What's inside

    Frequently Asked Questions

    Because the extraction is a pointer, not the paper. It normalizes formats, it can misread a table, and it has no evidence units at all in legacy records. Verify the number against the source, then cite the paper.

    That is a red flag, not a clean result. Either your question variants missed the contradictions or the search was biased. Run a variant phrased to find null results before concluding anything.

    Only if the connection worked. A connection error is not an empty evidence set — report the failure. If the search genuinely ran, a null result is still a finding: record it, check the DOI lookup for key papers, and note the coverage gap.