Research prompts that keep claims connected to evidence

Start with a question a source could answer

“Research onboarding” is too broad to evaluate. “Which steps in the supplied onboarding logs account for the longest waits?” establishes an evidence base and an observable result. It also keeps the model from replacing your problem with a generic overview.

State a time period, population, and scope when they matter. An answer about enterprise customers may not describe self-service users. A source from a previous product version may no longer describe the current workflow.

Choose a source mode

Supplied-source mode: the model works only from documents you provide. Ask it to distinguish what those documents say from its interpretation. Missing evidence should remain visible.

Browsing mode: use an AI tool that actually has web access. Ask for primary sources, dates, and source URLs. Requesting browsing in text does not enable a tool that the application does not have.

Planning mode: when neither evidence nor browsing is available, ask for a research plan and search questions. Do not label a plan as completed research.

A reusable claim-evidence prompt

Question: [one specific research question]
Scope: [population, period, and boundaries]
Evidence: [labeled documents or verified source material]

Use only the supplied evidence. Build a table with: claim, supporting source and excerpt, conflicting evidence, limitation, and verification needed.
Distinguish direct observations from interpretations. Do not treat a repeated claim as independent corroboration if the sources rely on the same underlying report.
Then write a synthesis of no more than 250 words. End with the most important unresolved question. If the evidence cannot answer the research question, say that plainly.

Giving documents labels such as “Interview A” and “Support log B” makes references easier to audit. Preserve those labels when splitting a long document into sections.

Worked example: a delay is not automatically a cause

Imagine a fictional onboarding log where seven customers waited after an identity-check step. A weak synthesis says identity checks caused seven delays. A careful synthesis says the waits occurred after that step and identifies what must be checked: missing customer documents, queue capacity, or a downstream dependency.

The prompt should request competing explanations and the evidence needed to distinguish them. It should not force a decisive conclusion when the available data supports only a narrower observation.

Review the claims that carry the conclusion

1. Open the cited source and locate the supporting passage.
2. Check that the passage supports the exact claim, not just the topic.
3. Check date, scope, and whether the source is primary or repeating another source.
4. Look for omitted qualifications or conflicting findings.
5. Revise the conclusion if the strongest evidence is weaker than the wording implies.

Ask for uncertainty in concrete terms: which value is unknown, what evidence is missing, and how the decision might change. A confidence percentage without calibration can add decoration instead of clarity.

Turn the synthesis into a useful next action

A good research deliverable can end with “We cannot yet distinguish A from B; collect this one observation next.” That is often more actionable than an exhaustive overview. Use the evaluation guide to check citation support and uncertainty handling across multiple cases.