The Source Contradiction Mapper: Turn Conflicting Research Into a Clear Picture of What's Actually Known
Why this prompt matters
Teams facing conflicting research typically do one of two costly things: cherry-pick the single study that confirms the decision they already wanted to make, or present an unhelpful it's complicated summary that gives leadership nothing to act on. Both lead to bad calls — either a six-figure tooling rollout based on cherry-picked data that later gets reversed, or weeks of stalled decision-making while stakeholders argue past each other because nobody has actually mapped where the sources agree, where they genuinely conflict, and why.
What we use it for
A product manager at a mid-size software company has been asked by leadership to recommend whether to expand AI coding assistant licenses from a 20-person pilot team to all 400 engineers. They have pulled five research reports and internal surveys that reach opposite conclusions about whether the tools actually help experienced developers, and need to brief the VP of Engineering by Friday without just picking whichever study confirms what leadership wants to hear.
Prompt
Role: Act as a rigorous research analyst whose specialty is synthesizing conflicting sources without flattening genuine disagreement into false consensus. Context: I am researching the following question: [YOUR RESEARCH QUESTION]. Below are summaries, abstracts, or excerpts from [NUMBER] sources, each labeled with a source name and date: [SOURCE 1 NAME, DATE]: [PASTE SUMMARY OR EXCERPT] [SOURCE 2 NAME, DATE]: [PASTE SUMMARY OR EXCERPT] [SOURCE 3 NAME, DATE]: [PASTE SUMMARY OR EXCERPT] [ADD MORE SOURCES AS NEEDED, 3-8 TOTAL WORKS BEST] Task: Read all sources carefully and produce a structured synthesis in four parts: 1. CONSENSUS — claims that at least 2/3 of sources agree on, with source citations for each claim. 2. DIRECT CONTRADICTIONS — specific pairs or groups of sources that disagree, stated precisely (not vaguely), with a note on whether the disagreement is about facts, methodology, definitions, or interpretation. 3. CONFIDENCE-GRADED SYNTHESIS — your best current-state summary of what is actually known, with each claim tagged [HIGH CONFIDENCE], [CONTESTED], or [SPECULATIVE] based on how many independent sources support it and how strong their evidence is. 4. OPEN QUESTIONS — what none of the sources address, or what would need to be true to resolve the contradictions in section 2. Constraints: - Do not average conflicting claims into a mushy middle position that no source actually holds — if Source A says X and Source B says the opposite of X, state that clearly, do not invent a compromise position. - Every claim must cite which source(s) support it. No unattributed claims. - If a source is clearly lower-quality (anecdotal, single case study, no methodology given) versus another that has clear methodology, note the quality difference — do not treat all sources as equally authoritative by default. - Flag your own uncertainty explicitly rather than projecting false confidence. Output format: Use the four numbered sections above as headers. Under each, use bullet points with inline source citations in parentheses. Keep the whole output under 600 words — the goal is clarity, not exhaustiveness.
Result
1. CONSENSUS
- AI coding assistants measurably speed up small, well-defined tasks: boilerplate generation, syntax lookup, and test scaffolding (TechCorp Internal Study 2026, DevSurvey Q2 2026, Stack Overflow Developer Report 2026).
- Adoption is now near-universal among professional developers — over 80% report using an AI coding tool at least weekly (DevSurvey Q2 2026, GitHub Octoverse 2026).
2. DIRECT CONTRADICTIONS
- Net productivity effect for experienced developers: METR's randomized controlled trial (July 2026) found experienced open-source contributors were 19% SLOWER on real tasks when using AI assistants, despite believing they were faster. TechCorp's internal study claims a 35% velocity increase for the same seniority band. This is a methodology conflict, not just a disagreement: METR used a blinded RCT with actual time tracking; TechCorp used self-reported estimates and ticket-closure counts, which the METR paper specifically flags as unreliable for this exact reason.
- Code review burden: DevSurvey reports 62% of senior engineers say AI-generated code increases their review workload; GitHub's Octoverse report claims review time per PR has stayed flat. These may not actually conflict — Octoverse measures aggregate time, DevSurvey measures perceived cognitive load, which are different metrics being talked past each other.
3. CONFIDENCE-GRADED SYNTHESIS
- [HIGH CONFIDENCE] AI tools speed up junior developers and well-scoped, low-ambiguity tasks.
- [CONTESTED] Whether AI tools speed up or slow down experienced developers on complex, real-world tasks — the only randomized controlled evidence (METR) says slower; self-reported industry surveys say faster, but self-report is a known-weak methodology for this specific question.
- [SPECULATIVE] Long-term skill atrophy effects — no source has data past 12 months of tool use.
4. OPEN QUESTIONS
- No source separates productivity effects by codebase size or legacy-code complexity — this is likely a major confound.
- None of the sources track outcomes past initial code merge (bug rates 3-6 months later are unmeasured).
- Resolving the TechCorp vs. METR contradiction would require TechCorp to re-run their study with blinded time tracking instead of self-report.
Most research synthesis prompts quietly do something dishonest: when sources disagree, they blend the disagreement into a soft, hedge-everything middle position that no actual source holds. The Source Contradiction Mapper is built to refuse that shortcut.
Why averaging disagreement is worse than useless
If Source A says a change increases output by 35% and Source B's controlled study says the same change decreases output by 19%, the honest synthesis is not their average. It's a flag: these sources disagree about the actual direction of the effect, and here's why — usually a methodology difference, not just noise. A model that smooths this into we're not sure but there might be a slight positive effect has thrown away the single most useful piece of information in the research: that the disagreement itself is diagnostic.
Why the four-section structure exists
Consensus comes first because it's the load-bearing foundation — what do most sources actually agree on, regardless of the noisier disputes layered on top. Direct Contradictions comes second and does the real work: it forces the model to name specific source pairs and specific claims, rather than gesturing vaguely at mixed findings. The confidence-graded synthesis section is where the prompt earns its keep for decision-makers — tagging claims HIGH CONFIDENCE, CONTESTED, or SPECULATIVE gives a reader something they can actually act on, instead of a wall of hedged prose. Open Questions closes the loop by naming what would need to be true to resolve the contradictions, which turns a one-time synthesis into a research roadmap.
Where this earns its keep
This prompt is built for exactly the situation most research write-ups fail at: you have several credible-looking sources that don't agree, and you need to brief someone who will make a real decision based on your summary. It works for competitive intelligence (comparing analyst reports on a market), academic literature reviews (comparing study findings on a research question), product decisions (comparing user research that points different directions), and policy analysis (comparing think-tank reports with different conclusions).
The one habit this prompt is designed to break
Confirmation bias in research synthesis rarely looks like outright dishonesty — it looks like unconsciously weighting the source that agrees with your prior belief a little more heavily, and skimming past the one that doesn't. Forcing an explicit Direct Contradictions section with named source pairs makes that kind of quiet cherry-picking much harder to do by accident, because the prompt requires you to state the disagreement plainly before you're allowed to move on to a synthesis.