Prompting for Research and Analysis

You're Using AI for Research the Wrong Way — and Getting Answers That Sound Right But Aren't.

There is a version of AI-assisted research that feels productive but is quietly unreliable: ask the model a factual question, receive a confident-sounding answer, and treat that answer as a reliable source. The output can be accurate. It can also be confidently wrong. And without a clear strategy for distinguishing between the two, you are relying on luck.

The problem is not that AI is bad at research — it is that most people prompt for retrieval (give me the answer) instead of analysis (help me think through this). Retrieval prompts put the AI in a position where it has to be right about facts. Analysis prompts put the AI in a position where it helps you reason — and your job is to verify the facts it surfaces, not to accept them by default.

Retrieval vs. Analysis: The Fundamental Difference

Retrieval prompts ask the model to produce facts: "What was the market cap of the 10 largest US banks in 2023?" This is a task where the model can be confidently wrong, and where you have no easy way to tell the difference between a correct answer and a plausible-sounding hallucination without going to the source.

Analysis prompts ask the model to reason: "What factors typically determine whether a regional bank grows its market share during periods of rising interest rates, and what would you expect to be most important for a community bank with heavy mortgage exposure?" This is a task where the model's reasoning can be evaluated and pushed on, and where you are not depending on its factual memory being correct.

The practical rule: use AI for analysis, synthesis, and structuring your thinking. Use primary sources for facts you will act on.

The Steelman Prompt

One of the most useful research prompts in an intermediate's toolkit is the steelman request. Instead of asking the model to take a position, ask it to make the strongest possible case for a view you are evaluating — especially one you are inclined to dismiss.

Position-seeking: "Is decentralized finance actually a viable alternative to traditional banking?"

Steelman: "Make the strongest possible case for decentralized finance as a viable alternative to traditional banking for retail users. Assume a sophisticated skeptical audience. Do not hedge — argue the strongest version of the case."

The steelman forces the model to surface the best arguments for a position, which sharpens your thinking whether you agree with them or not. Follow with a counterargument request for the full picture.

The Breakdown Prompt

For complex topics you need to understand quickly, the breakdown prompt structures unfamiliar territory before you dive in.

Standard: "Explain quantum computing to me."

Breakdown: "I need to understand quantum computing well enough to evaluate a vendor's quantum-readiness claims in a procurement conversation. Break down the topic into: (1) the core concepts I need to understand, (2) the claims vendors typically make that are currently realistic vs. overhyped, and (3) three questions I should ask in the conversation to test their credibility. I have 10 minutes to read this."

The breakdown prompt defines the purpose (evaluate vendor claims), the format (structured sections), and the constraint (10 minutes). It produces a targeted primer rather than a Wikipedia summary.

The Synthesis Prompt

When you have gathered material from multiple sources, the synthesis prompt asks the AI to do the intellectual work of integration — identifying patterns, contradictions, and implications across the material.

Basic: "Here are three articles about remote work productivity research. Summarize them."

Synthesis: "Here are three articles about remote work productivity research. Compare their main findings: where do they agree, where do they disagree, and what would explain the contradictions? Then identify the single most important implication for a manager who is deciding whether to require in-office days."

The Assumption Audit

One of the most underused research prompts: asking the model to surface the hidden assumptions in your question. This is particularly useful early in a research project when you may not know what you do not know.

Assumption audit: "I am about to research whether my company should build its own AI tools or buy off-the-shelf solutions. Before I start, what assumptions am I already making in how I framed this question, and what alternative framings might be more accurate or useful?"

Verifying What You Get

Any factual claim in an AI-generated analysis that you will act on needs verification against a primary source. The AI should help you find the claim and frame the question; the source provides the ground truth. Build this into your research workflow as a standard step, not an exception. AI is most reliable as a thinking partner and synthesis engine — not as a fact database that you can trust without checking.