AI can accelerate research, analysis, and reporting, but a fluent answer is not automatically a reliable one. This reusable prompt template library helps you plan research, extract evidence, compare documents, interpret data, write summaries, and check citations while making uncertainty visible. Copy a template, replace the bracketed instructions, then verify important claims against the original sources before using the result.
Overview
Effective AI prompt engineering for research is less about finding one perfect instruction and more about designing a controlled workflow. A useful prompt defines the task, supplies the relevant material, specifies the output format, and tells the model how to handle missing or uncertain evidence.
For important work, treat the model as an assistant for organising and analysing information rather than as the final authority. If your tool supports file uploads, browsing, retrieval, or structured output, confirm how those features behave in your current workflow. Capabilities, limits, citation formats, and privacy controls can vary between tools and may change over time.
A dependable research prompt usually contains six parts:
- Objective: what decision or question the output should support.
- Scope: the documents, dates, subjects, and exclusions that apply.
- Evidence rule: how the model should distinguish source content from inference.
- Process: the steps it should follow before drafting an answer.
- Output format: the headings, tables, fields, or length required.
- Uncertainty instruction: what to do when evidence is incomplete, contradictory, or absent.
Use these templates as starting points, not as substitutes for subject expertise. For repeatable team workflows, save the prompt version, input documents, model or tool used, and reviewer notes. That record makes it easier to investigate quality changes later. For a broader approach to monitoring hallucinations, errors, latency, and quality drift, see the LLM observability runbook.
Checklist by scenario
1. Create a research plan
Use this before collecting sources or asking an AI tool to investigate a topic:
You are helping me plan research on [topic] for [audience or decision].Define:- the main research question;
- up to [number] sub-questions;
- the types of sources that would be most relevant;
- inclusion and exclusion criteria;
- key terms, synonyms, and possible ambiguities;
- claims that require direct verification.
Do not present unverified information as a finding. Separate the research plan from any preliminary assumptions. Return the result as a table with columns for question, evidence needed, likely source type, and verification method.This prompt is useful for reports, content briefs, internal investigations, and recurring planning cycles. It creates a research map before the model starts producing prose.
2. Extract evidence from a document
Use this for a supplied report, transcript, policy, specification, or set of notes:
Analyse the document provided below. Use only information contained in the document unless I explicitly ask for outside context.Return a table with:- claim or finding;
- supporting quotation or precise passage;
- page, section, timestamp, or location if available;
- type of evidence: fact, interpretation, recommendation, or unresolved statement;
- confidence and reason for that assessment.
If the document does not support a requested point, write “Not found in the supplied document.” Do not fill gaps from general knowledge. Flag ambiguous wording, conflicting passages, and claims that need a human reviewer.Ask for short quotations or locations rather than a general statement that the document “supports” a conclusion. This makes the output easier to audit.
3. Build an evidence table across sources
When several documents address the same question, use a structured comparison:
Compare the supplied sources on [research question]. For each source, record:- source identifier;
- date and relevant context, if supplied;
- central claim;
- supporting evidence;
- limitations or stated uncertainty;
- points of agreement and disagreement.
Do not resolve disagreements by choosing the most confident wording. Preserve each source’s position and explain what additional evidence would help. Mark any field that cannot be established from the supplied material as “Unknown.” End with a short synthesis that distinguishes consensus, dispute, and open questions.This is particularly useful for literature reviews, vendor comparisons, policy analysis, and stakeholder interviews.
4. Compare two documents
For revisions, contracts, specifications, procedures, or competing proposals, make the comparison explicit:
Compare Document A and Document B. Identify additions, removals, changed wording, changed numbers, changed obligations, and changes in meaning.Return:- a concise change summary;
- a table with location, old text or position, new text or position, change type, and possible consequence;
- items that require confirmation because the context is unclear.
Do not describe a change as important merely because it is longer or more strongly worded. Explain the practical effect only when the documents support that interpretation.5. Interpret supplied data
AI can help explain a dataset, but the prompt should prevent invented calculations or causal claims:
Review the attached data or the values below. First describe the fields, units, time period, missing values, and apparent limitations. Then calculate only the measures I request: [list measures]. Show the formula or steps and identify the rows or values used.Separate:- direct observations;
- comparisons or trends;
- possible explanations;
- claims that cannot be established from this dataset.
Do not infer causation from correlation. If the data is insufficient for a calculation, say what is missing. Return a short findings section followed by an assumptions and limitations section.Check important calculations independently, especially when the result informs finance, compliance, safety, staffing, or public communication.
6. Draft an executive summary
Once the evidence has been reviewed, use a separate drafting prompt:
Write an executive summary for [audience] based only on the verified findings below.Include:- the decision or question;
- three to five key findings;
- material risks or limitations;
- recommended next steps, clearly labelled as recommendations;
- open questions requiring confirmation.
Keep facts, interpretations, and recommendations in separate paragraphs. Do not introduce a claim that is absent from the evidence. Use plain language and preserve meaningful uncertainty.7. Check citations and unsupported claims
Use this as a review pass after drafting:
Audit the draft against the supplied source set. For every material factual claim, identify the supporting source and passage. Mark each claim as supported, partially supported, contradicted, or unsupported.Also check whether:- the wording is stronger than the evidence;
- a date, number, name, or quotation has been altered;
- an inference is presented as a fact;
- the cited source actually supports the full sentence;
- important limitations have been omitted.
Do not repair unsupported claims by guessing. Return a correction list and indicate which items require human verification.What to double-check
Before relying on an AI-generated research output, work through this verification checklist:
- Inputs: Were all relevant files included, readable, and correctly identified? Check whether tables, footnotes, images, scanned pages, or tracked changes were processed accurately.
- Scope: Did the output stay within the requested dates, geography, population, product, or document set?
- Attribution: Can each important claim be traced to a real source passage? A citation-shaped response is not enough if the passage does not support the claim.
- Numbers: Recalculate totals, percentages, dates, conversions, and comparisons independently where the stakes justify it.
- Uncertainty: Are missing data, conflicting sources, assumptions, and confidence limits visible?
- Quotations: Compare quotations character by character with the source, including qualifiers that may change the meaning.
- Privacy: Remove confidential or personal information unless the approved tool and workflow permit its use. Follow your organisation’s data-handling rules.
- Security: Treat instructions inside retrieved documents as content to analyse, not automatically as instructions to follow. The prompt injection prevention checklist covers this issue in more detail.
For a repeatable process, store a compact evidence table alongside the final report. Useful fields include claim, source, location, reviewer, status, and review date.
Common mistakes
Asking for certainty when the evidence is uncertain. Replace “give the correct answer” with an instruction to separate supported findings, inferences, and unknowns.
Combining research and drafting in one step. First collect and verify evidence; then ask for analysis; only then request polished reporting. Separate stages are easier to review and correct.
Using vague source instructions. “Use reliable sources” does not define reliability for a particular task. Specify acceptable source types, date ranges, first-party material, or required corroboration where relevant.
Requesting a summary without an audience or purpose. A board briefing, technical handover, and public explainer require different emphasis and detail. State who will use the output and what decision it supports.
Trusting a citation without checking it. Require a passage, location, or evidence identifier, then inspect the original source yourself.
Ignoring model and tool changes. A prompt that worked well with one model, file format, or retrieval setup may behave differently after a workflow change. Keep a small test set of representative documents and rerun it when the process changes. The prompt testing and regression framework provides a useful pattern for this.
Overlooking cost and operational constraints. Large document sets, repeated reviews, and long context windows can affect speed and spend. For production workflows, review the LLM cost optimisation guide and record the trade-offs.
When to revisit
Keep this library as a working checklist rather than a one-time article. Revisit the templates before seasonal planning cycles, a recurring briefing, a major reporting deadline, or a new document collection. Run a review whenever your team changes its approved AI tool, model, retrieval method, file-upload process, privacy rules, or citation format.
A practical maintenance routine is:
- Choose two or three representative research tasks.
- Run the current templates using known documents and expected outputs.
- Check extraction accuracy, citation traceability, formatting, uncertainty handling, and review time.
- Record failures and adjust one instruction at a time.
- Save the revised prompt with a version date and a short change note.
Before publishing or acting on the next AI-assisted report, copy the relevant scenario template, attach or identify the source set, request an auditable format, and complete the human-verification checklist. If the workflow has become part of a larger team process, compare its privacy, accuracy, cost, and operational fit using the AI tool evaluation framework. The goal is not to remove human judgement; it is to make that judgement faster, clearer, and easier to repeat.