Research Agent Brief: Source Discovery, Evidence Review, and Citation Table

Directs an AI agent to plan source discovery, grade evidence, build a citation table, and support a decision without claiming live browsing occurred.

You get: Search plan table, citation table and synthesis by question

More like this

The prompt

research-agent-brief-source-discovery-evidence-review-and-citation-table.md
Role: You are a research agent that plans and structures evidence-based investigations.Context: I need to decide {decision} about {topic} for {audience}. You may not have live web access. Do not claim you searched, opened, or verified any source unless I paste the content or your tool explicitly returns it.
Task: Produce a research plan, then evaluate the evidence I supply, then give a decision-support synthesis.Constraints: Ask for missing inputs first. Keep assumptions separate from evidence. Do not fabricate citations, quotes, statistics, URLs, or authors. If you recall a source from memory, label it 'unverified recall - confirm before use'. No guaranteed conclusions.

Inputs:
- Decision and deadline: {decision}
- Topic and scope (include exclusions): {topic}
- Audience and their risk tolerance: {audience}
- Sources I will paste (links, excerpts, notes): {pasted_sources}
- Preferred source types and date range: {source_criteria}
- Key questions to answer: {research_questions}

Workflow:
1. Ask for missing inputs.
2. Source discovery plan: give 8-12 search queries, grouped by research question, plus source types to prioritize (primary data, regulators, peer-reviewed, vendor docs, reputable press) and where each is usually found. State that I must run these searches.
3. Evidence review: for each pasted source, record claim, method or basis, date, potential bias, and strength (High/Medium/Low) with reasons.
4. Build the citation table from pasted sources only.
5. Synthesis: findings per research question, agreements, conflicts, gaps.
6. Decision support: options, what the evidence supports, confidence level, and what would change the recommendation.

Output format:
1. Assumptions and missing information
2. Search plan table: Question | Query | Source type | Why it matters
3. Citation table: ID | Source | Date | Claim | Basis | Bias risk | Strength
4. Synthesis by question, citing IDs
5. Decision brief: Recommendation | Confidence | Key risks | Evidence gaps
6. Next actions (verification steps, experts to consult)

Acceptance criteria:
- Every factual claim in the synthesis cites a table ID or is marked as an assumption.
- No source appears that I did not provide, except labeled unverified recall.
- Conflicts and gaps are stated, not smoothed over.
- The recommendation includes what would change it.

Quality checks:
- Re-read the synthesis and remove any claim without an ID.
- Check that the confidence level matches the evidence strength.
- Confirm that the search plan is not described as completed.

Platform notes: In ChatGPT or Gemini with browsing enabled, tell me which results came from the tool and include the links it returns. In Claude, keep the citation table in Markdown and put the decision brief in <decision> tags. Without browsing, only work with pasted material.

Example output

AI-written example

1. Assumptions and missing information

  • Assumption: the deadline is 30 June and the audience (support leadership) is moderately risk-averse about customer data.
  • Missing: the tool's data-retention terms, the pilot's sample size, and the budget ceiling.
  • No searches have been run. The plan below must be run by you.

2. Search plan (not yet executed)

QuestionQuerySource typeWhy it matters
Q1 Time saved?"AI meeting summarization time savings controlled study"Peer-reviewedIndependent effect size
Q1"call center AI note-taking productivity evaluation"Reputable press, journalsComparable setting
Q2 Privacy risk?"AI transcription tool data retention GDPR support calls"RegulatorsCompliance exposure
Q2"[vendor] data processing addendum subprocessors"Vendor docsContract terms
Q2"AI transcription data breach incident"Reputable pressReal-world failures
Q3 Accuracy?"speech-to-text word error rate accented speech benchmark"Peer-reviewedError rates on your callers
Q3"AI summary hallucination rate customer service"Primary dataQuality risk
Q3"[vendor] independent audit accuracy"Vendor docs, third-party reviewsCheck marketing claims

3. Citation table (pasted sources only)

IDSourceDateClaimBasisBias riskStrength
S1Vendor case study (pasted)Mar 2024Agents save 6 min per callCustomer self-reportHigh: vendor-authoredLow
S2Internal pilot notes (pasted)May 202411 agents saved about 3 min per call; 4 summaries had errorsUnblinded 2-week trialMedium: small sample, enthusiastic volunteersMedium
S3Trade survey excerpt (pasted)Jan 202458% of support teams cite privacy as the main barrierSurvey, n not statedMedium: sample unclearLow

4. Synthesis

  • Q1 Time saved: S1 and S2 agree on a positive direction but differ in size (6 vs 3 min). Neither is independent (S1, S2).
  • Q2 Privacy: S3 suggests privacy is a common concern, but nothing pasted covers this vendor's retention terms. Gap.
  • Q3 Accuracy: S2 reports 4 flawed summaries, but the total number reviewed is not given (S2). Gap.
  • Conflict: the S1 vs S2 time-savings estimates.

5. Decision brief

<decision>
Recommendation: Proceed only with an expanded, 30-agent pilot with human review of summaries. Do not commit to a full rollout yet.
Confidence: Low to Medium. Evidence is directionally positive (S2) but small and non-independent.
Key risks: Unreviewed contract terms for call data; summary errors reaching customer records (S2); overstated vendor savings (S1).
Evidence gaps: Retention and subprocessor terms; error rate denominator; independent benchmarks.
What would change this: Unaccept

Generated from this prompt by AI. Your results will vary.

Tips and common edits

Run in two passes: first to get the search plan, then paste your collected sources for the evidence review. Verify every citation yourself before publishing or acting.

Questions about this prompt

More in Agentic AI

Parallel Research Fan-Out and Fan-In with Citation Table and Decision Memo

Splits a research question across parallel research agents, defines source discovery and evidence review steps, merges findings into a citation table, and produces a decision memo without claiming live browsing.

Competitor Landscape Research Plan with Source Discovery and Citation Table

Builds a source discovery plan, evidence review, citation table, and decision-ready synthesis for a competitor or market question, without pretending live browsing occurred.

Competitor UX Pattern Research Plan with Evidence Table and Design Recommendation

Builds a source discovery plan, evidence log, citation table, and synthesis for comparing UX patterns across competitors, without claiming live browsing.

Competitor and Audience Evidence Review with Citation Table and Go/No-Go Brief

A research workflow that plans online source discovery, scores evidence, builds a citation table, and ends with a decision brief, without pretending live browsing happened.

Competitor Landscape Scan with Source-Backed Citation Table

Runs a structured competitor scan that starts with search queries, ranks sources by reliability, and ends with a cited comparison table and a positioning recommendation.

Evidence Review Plan for a Classroom Teaching Strategy Decision

Guides source discovery, evidence review, a citation table, synthesis, and a recommendation on whether to adopt a teaching strategy. It does not claim live browsing and marks every unverified item for you to check.

Productivity Tool Selection: Source Discovery, Evidence Review, and Citation Table

A research workflow to choose between candidate productivity or automation tools. It plans online source discovery, structures evidence review in a citation table, and ends in a decision memo without claiming live browsing took place.

Audience Research Source Discovery and Evidence Synthesis for Segment Prioritization

Guides you to plan source discovery, review the evidence you collect, build a citation table, and synthesize findings into a segment prioritization decision, without pretending the model browsed the web.

20-Case Agent Eval Set with Pass/Fail Rubric for Normal, Edge and Unsafe Requests

Builds 20 test cases for your agent's task, with expected behavior, failure signals and a pass/fail rubric covering normal, edge and unsafe requests.

Agent Tool Names, Descriptions and JSON Schemas from API Docs

Turns pasted API docs or a function list into agent-ready tool names, descriptions, JSON input schemas and when-to-use and when-not-to-use notes.

Agent Workflow Decomposition with Autonomy Levels and Stop Rules

Turns a manual process into a step-by-step agent plan with tools, approval gates, retry limits, and a build/pilot/skip recommendation.

AI Paper Explainer with Evidence Table and Practical Takeaways

Turns a pasted AI paper abstract or text into a plain-language explanation of what is new, how it was tested, its limits and what to do with it.