Claude Research Prompts That Preserve Evidence

Search-design prompt

Use this before touching a database: "I am reviewing [topic] in [population/corpus] from [date range]. Generate candidate search terms, controlled-vocabulary candidates, synonyms, adjacent terms, and likely false positives. Return a table with term, why it matters, where to test it, and what it might accidentally retrieve."

Follow-up: "Convert the strongest term groups into database-specific query drafts for PubMed, OpenAlex, and arXiv. Do not claim these searches have been run."

Paper-triage prompt

Use this after exporting records: "Classify each record against these inclusion criteria: [criteria]. Use labels include, exclude, maybe, wrong corpus. Give one abstract-grounded reason and quote no more than a short phrase from the abstract as the anchor. If the abstract is insufficient, label maybe."

This prompt makes uncertainty visible. The "maybe" pile is where expert time should go.

Evidence-extraction prompt

Use this with included full text: "Extract the following fields only from the provided source: study design, sample/corpus, method, intervention/exposure, comparator, outcome, effect or finding, limitation, funding/conflict, and passage marker. Use not reported when absent. Do not infer values from background knowledge."

Follow-up: "Which extracted fields are weakest, ambiguous, or dependent on a figure/table I should inspect manually?"

Synthesis and contradiction prompt

Use this after the evidence table is locked: "Draft a synthesis from the evidence table only. For each paragraph, list the source IDs that support it. Add a separate section called Claims to weaken, naming the evidence that most complicates each paragraph."

This turns Claude into an argument drafter and checker, not a replacement for literature judgment.

Reproducibility prompt

For computational papers or Claude Science outputs, use: "Create a reproducibility checklist from this method and artifact history. Include data availability, code availability, environment, package versions, random seeds, hardware or compute assumptions, external tools, and any step that cannot be reproduced from the record."

The output should become a notebook entry and, when appropriate, a reviewer question.

Workflow checklist

1. Paste the relevant template.
2. Replace bracketed placeholders with protocol details.
3. Require fields rather than prose.
4. Require "not reported" for missing information.
5. Run a contradiction pass.
6. Save the prompt and output in the research notebook.

Researcher FAQ