Research Prompts for AI: Evidence You Can Trace

Good research prompts require an answerable question, a defined source set, traceable evidence, and an explicit account of what remains unknown. Use separate prompts to plan the search, extract findings, and write the conclusion. That separation gives you a chance to catch a weak source before it becomes a polished recommendation.

fenbs can keep the reusable prompts, source links, and open questions with the project, so the next connected assistant can continue from a checked record. Keep research documents in the systems approved for them, and save the instructions and work status you need to reuse.

How do you turn a broad topic into a research question?

Begin with the decision the research should inform. “Research remote work” is too broad to tell you when you have enough evidence. “What evidence would help a small US design studio choose a meeting schedule?” has a reader, a decision, and a practical stopping point. This is a hypothetical example, not a finding about remote work.

Help scope research for this decision: [decision].
Audience: [reader]. Setting: [US location, industry, population].
Return one main question, three subquestions, key definitions, and evidence that would change the decision.
Separate factual questions from preferences a person must settle.
Propose inclusion and exclusion criteria, a search date range with reasons, and a stopping condition.
Do not answer from memory or invent sources. Ask about any missing detail that changes the scope.

Review the proposed limits before collecting evidence. If the studio needs scheduling advice for a small creative team, a study of an unrelated industry may offer context without directly answering its question. Agree on what counts as relevant and keep that criterion stable unless a person approves a reason to change it.

What should a source-finding prompt ask for?

Request primary material: the original study, official dataset, agency publication, or vendor documentation behind a claim. A review can help you discover that material, but distinguish its interpretation from the underlying evidence. Include opposing findings and missing evidence in the assignment so the search does not merely confirm your preferred answer.

Find primary sources for [approved question and scope].
First state whether you can search and open pages in this session. If you cannot, provide search queries only.
For each candidate, list title, publisher or authors, publication date if shown, URL, source type, and the subquestion it may answer.
Label each source: opened, metadata only, or inaccessible.
Include contradictory evidence and explain coverage gaps. Do not treat a search snippet as a source you read.

For biomedical literature, the PubMed user guide↗ (opens in a new tab) explains searches, filters, and citation lookup. Use the database’s own search tools to inspect what the assistant found; an AI-generated reading list is a set of leads. Do not call it an exhaustive literature search unless you have carried out and documented the appropriate method.

Separate guides cover AI search tools, research workflows, market research, academic research tools, and legal research when you need a tool choice or a specialized method.

How can you extract findings without overstating them?

Give each source a short label and preserve the location of every important passage. Ask for the population, method, comparison, and limitations alongside the result. Without those fields, a narrow observation can become a universal statement as it moves through successive summaries.

Read source [label]: [approved text or accessible source].
Answer only from this source.
Return: question studied; population or setting; method; comparison; main findings; limitations; funding or conflicts if stated; and unanswered questions.
For each finding, give a short supporting passage and its page, table, or section.
Separate what the authors found from their interpretation.
If you have only an abstract, label the card abstract-only and do not infer details from the unavailable full text.

Check quotations against the original and keep excerpts limited to what your work needs. Read the surrounding passage, including qualifications and table notes. If a numerical finding matters to the decision, verify the denominator, units, time period, and comparison yourself. A correct number attached to the wrong group still produces a wrong conclusion.

Save the reusable extraction instructions and their required checks in fenbs Skills and prompts. A future assistant should know that an abstract-only card remains incomplete, rather than treating every saved note as equally strong evidence.

How do you compare sources that disagree?

First check whether the sources answer the same question. Different populations, definitions, and dates may explain an apparent disagreement. Ask for a comparison organized by those differences before asking which conclusion to trust. Avoid resolving disagreement by counting how many pages repeat each claim.

Compare these checked evidence cards: [paste].
Create a table with question, source label, population, method, outcome definition, finding, and limitations.
Identify genuine contradictions separately from differences in scope.
List shared underlying datasets so repeated reporting is not counted as independent evidence.
For each disputed point, state what additional evidence would help.
Do not choose a winner by source count or invent a confidence percentage.

Look for repeated use of the same underlying study. Several articles describing it do not give you several independent tests. Also distinguish “no evidence found within this search” from “evidence proves there is no effect.” The first describes your search; the second requires a much stronger evidentiary basis.

How should you verify references and quotations?

Open every reference you intend to cite. Match title, authors, date, and publication to the actual record, then check the relevant passage. A real publication can still be cited for something it never says. Verify both identity and support, and mark anything you could not access.

Crossref’s metadata documentation↗ (opens in a new tab) describes records deposited by publishers and trusted sources. Those records can help match a scholarly reference to its identifier and publication details. Matching metadata confirms a record exists; it does not establish that its findings support your claim or that the work has no limitations.

• Citation identity: does the reference match the original publication?
• Claim support: does the cited passage actually support your sentence?
• Scope: have you preserved the population, conditions, and uncertainty?
• Current status: check the publisher record for relevant corrections or retractions.
• Access: distinguish a document you read from a citation or abstract you could only identify.

Anthropic’s hallucination guidance↗ (opens in a new tab) recommends source grounding and permission to acknowledge uncertainty. Those are useful prompt requirements, but a model reviewing its own answer is not independent confirmation. Retain a human check of the original source.

What should a synthesis prompt produce?

Request a conclusion proportional to the evidence. Give the assistant only checked evidence cards and ask it to connect each conclusion to specific sources. Keep observations, interpretation, and recommendations visibly separate. A recommendation can involve a business preference even when every observation is accurate.

Answer [question] using only these checked evidence cards: [paste].
Structure: short answer; supported findings with source labels; disagreements; evidence gaps; and practical options.
Separate observations from interpretation and recommendations.
Preserve limitations and abstract-only labels.
Do not add new factual claims or sources.
End with what a person must decide and what further evidence could change the answer.

For ChatGPT or Claude, start by stating whether this is a search assignment or an analysis of supplied material. Use the same evidence fields in either app and confirm actual tool access in the session. Portable wording does not guarantee the same search coverage, document access, or answer quality.

The general ChatGPT prompts for work collection includes shorter research starters; system prompt examples cover standing instructions. An outside work record can preserve checked progress, but evidence still needs a traceable source.

How to do this with fenbs

1. Create a free account and add a project describing the research question and decision it informs.
2. Add the prompts you have tested to Skills and prompts. Include their required inputs, output fields, and source-checking steps.
3. Record important source URLs in Links and listings, and add tasks for inaccessible evidence, disputed findings, or checks still owed.
4. Connect a supported assistant using its setup guide. Ask it to read the project context and relevant prompts before working from approved source material.
5. Save progress before stopping: what was checked, what remains uncertain, and the next source or question to examine. Keep confidential documents in their approved system.

fenbs is free for one person, with no credit card. It is useful for keeping reusable instructions and research progress across sessions; it is not a scholarly database or a replacement for a reference manager. Explore the AI prompts collection when you need a different kind of assignment, and start here with one question whose sources you can check.