Research Prompts for Source-Grounded Answers, Fact Checking, and Evidence Synthesis

Overview

AI can help researchers move quickly, but speed creates a risk: a fluent answer may combine verified information, incomplete context, reasonable inferences, and unsupported claims. Prompt engineering cannot replace source review. It can, however, make the research workflow more explicit and easier to audit.

A dependable process separates research into stages:

1. Define the question: Set the scope, audience, date range, geography, and required level of evidence.
2. Expand the query: Turn the original question into related search terms, competing explanations, and missing subquestions.
3. Collect sources: Gather relevant pages, papers, reports, transcripts, datasets, or uploaded files.
4. Extract evidence: Record what each source actually says, including limitations and publication details.
5. Map claims to evidence: Connect every important statement to a supporting passage or mark it as unresolved.
6. Analyze disagreement: Explain whether sources differ because of definitions, methods, dates, samples, or interpretation.
7. Synthesize carefully: Write a conclusion that reflects the strength and limits of the evidence.

This staged approach works for an article brief, a literature review, competitive research, product research, or an internal report. It also makes prompts easier to test because each prompt has one primary job. For structured records, see the guide to structured output prompts and JSON Schema. For long source files, the guidance on prompting for long documents and large context windows can help you divide the material into manageable passes.

Checklist by scenario

1. Query expansion and research planning

Use this prompt before browsing or asking a model to summarize sources:

Act as a research planning assistant. Do not answer the topic yet.

Research question: [insert question]
Purpose: [report, article, literature review, decision, or other]
Audience: [describe audience]
Scope limits: [date range, geography, industry, language, or exclusions]

Create:
1. A precise restatement of the question.
2. The main concepts that need definition.
3. Ten search queries covering direct evidence, opposing views, historical context, and likely blind spots.
4. The types of sources that would be most useful for each subquestion.
5. Questions that cannot be answered reliably without additional evidence.

Label suggestions as research leads, not verified facts.

Review the suggested queries before using them. Remove terms that introduce an assumption, and add the names of relevant methods, populations, products, locations, or time periods when precision matters.

2. Source extraction

When you have a source, ask the model to extract evidence rather than provide a general summary:

Analyze only the source provided below. Do not add outside information.

For each relevant passage, return:
- source identifier
- section, page, paragraph, or timestamp if available
- exact claim or a faithful short quotation
- evidence type: finding, definition, example, opinion, forecast, or limitation
- entities, dates, numbers, and conditions mentioned
- confidence that the passage supports the stated claim

If the source does not contain the requested information, write: Not found in source.
Separate direct statements from your interpretation.

Source: [paste text or attach file]

For a web-based workflow, preserve the URL and access date in your own notes. For uploaded files, retain the file name and page or section reference. Model-provided citations should be treated as leads until you confirm that the cited passage exists and supports the claim.

3. Claim-to-citation mapping

This prompt turns extracted material into an audit trail:

Using only the source notes below, build a claim-to-evidence table.

Columns:
1. claim ID
2. proposed claim
3. supporting source ID
4. exact passage or location
5. support level: direct, partial, contextual, or unsupported
6. important qualification
7. wording that would avoid overstating the evidence

Do not create a citation for a claim that is not supported. If several claims are combined in one sentence, split them into separate rows.

Source notes: [insert notes]

A useful evidence table can be maintained in a spreadsheet or database with these columns: Claim, Source, Location, Evidence, Publication date, Scope, Limitations, Status, and Reviewer note. Status values such as unreviewed, verified, disputed, and unresolved make the workflow easier to manage.

4. Disagreement analysis

Conflicting sources do not always represent a simple true-versus-false dispute. Ask:

Compare the sources below without selecting a winner prematurely.

For each disagreement, identify:
- the exact propositions that differ
- whether the difference is factual, definitional, methodological, temporal, geographic, or interpretive
- the evidence each source uses
- assumptions or limitations
- what additional evidence could resolve the disagreement

End with a cautious assessment of what can be stated confidently, what remains contested, and what should not be inferred.

Sources: [insert source notes]

5. Final synthesis

Use a separate synthesis prompt after the evidence table is reviewed:

Write a source-grounded synthesis from the verified evidence below.

Requirements:
- distinguish evidence from interpretation
- cite the source ID after each material claim
- preserve important dates, scope limits, and uncertainty
- do not fill gaps with general knowledge
- identify meaningful disagreement
- include a short section titled Evidence gaps
- if a conclusion is an inference, label it as an inference
- if evidence is insufficient, say so plainly

Audience: [describe audience]
Desired format: [brief, report, article section, or literature review]
Verified evidence: [insert evidence table]

What to double-check

• Source identity: Confirm the author, publisher, document title, URL, version, and publication or revision date.
• Passage accuracy: Open the source and check that the cited wording, page, section, or timestamp is correct.
• Claim scope: Make sure a narrow finding has not been rewritten as a universal statement.
• Time sensitivity: Check whether the source is still appropriate for the question, especially for products, regulations, market conditions, or software tools.
• Numbers and units: Verify calculations, currencies, denominators, percentages, dates, and whether figures describe a sample or a whole population.
• Primary evidence: When possible, follow summaries back to the original paper, filing, dataset, documentation, interview, or official record.
• Source independence: Several pages may repeat the same original claim. Count the underlying evidence, not just the number of pages.
• Model boundaries: Confirm which sources the model could actually access. A browsing or file-analysis feature does not remove the need for human verification.

If your workflow uses a model to return records, require consistent fields and reject incomplete rows. A structured response is easier to review, but formatting is not proof of accuracy. You can develop a repeatable review set with the prompt testing framework and add research-specific cases such as missing citations, unsupported conclusions, and conflicting dates.

Common mistakes

1. Asking for a fully researched answer in one prompt: This hides the difference between discovery, extraction, and verification. Use prompt chaining so each stage produces an inspectable output.
2. Requesting citations without supplying a source policy: State whether the model may use only supplied files, approved domains, or a defined source set.
3. Confusing a citation with support: A link can be relevant yet fail to support the exact sentence. Check the passage and reduce the claim when support is partial.
4. Allowing the model to complete missing information: Include explicit instructions such as Not found in source and unresolved rather than encouraging a complete-sounding answer.
5. Mixing notes and conclusions: Keep raw extraction, verified evidence, interpretation, and final prose in separate fields or documents.
6. Ignoring definitions: Terms that appear similar may refer to different populations, time periods, methods, or outcomes. Define them before comparing sources.
7. Failing to record changes: Save the prompt version, source set, model or tool used, date of analysis, and reviewer decisions so the result can be reproduced.

When to revisit

Revisit the workflow whenever the research inputs change, not only when the final article is due. Run a fresh source check before seasonal planning cycles, publication, a major business decision, or any update involving rapidly changing tools and documentation. Recheck the source set when a document is revised, a URL changes, a new primary source appears, or two credible sources begin to disagree.

Tool capabilities also change. Browsing behavior, file limits, citation displays, structured output support, and research features may differ across models and product versions. Maintain a short tool note in your research record: what the tool was allowed to access, which files were provided, how citations were returned, and which checks were performed manually. Do not assume that a prompt that worked in one environment will behave identically in another.

Before acting on a report, use this final checklist:

• Is the research question specific enough to test?
• Are the key terms and scope defined?
• Did query expansion include opposing views and blind spots?
• Does every material claim have a verified supporting passage?
• Are uncertainty, disagreement, and evidence gaps visible?
• Were dates, figures, links, and source locations checked?
• Is the final wording no stronger than the evidence?
• Have the prompt, source set, tool, and review date been recorded?

Keep the evidence table with the finished report. That small habit turns AI research prompts from one-off instructions into a repeatable AI development workflow for fact checking, source-grounded content, and evidence synthesis.