AI@GSB: Advanced AI Prompting for Research with Diego Oppenheimer
Stanford Graduate School of Business
Research by claims 0:05
Diego Oppenheimer says he uses AI for market research by making claims and then trying to break them. A generic report is not useful, because anyone can get the same consensus view from a model in about an hour. He wants to find what others are missing. For this session, he uses AI tooling for physical and hardware engineering, where he sees a possible shift like the one software engineering went through.
Scoping the investigation 9:01
He begins by having the model question him about the space, his rough hypothesis, time budget, and what he already knows. That keeps the model from leading and helps him build context from his own answers. From that first pass, he gets a claim ledger with items to test, such as hardware teams lacking a system of record, buyers paying about $200 per engineer each month, and a market size in the $5 billion to $10 billion range. He says he is not trying to outsource judgment, only to narrow the search.
Consensus Before Research 15:30
Before you start research, write down what everyone already believes. Treat that as the null hypothesis. If a claim takes more than a few minutes to pin down, treat it as straw man material and move on. The point is to turn broad talk into claims you can test, like a market being smaller than people say, rather than slogans that sound true but cannot be checked.
Track Money and Boundaries 23:00
Then trace how money moves through the market. Ask where a dollar of customer spend starts, which budgets and buyers touch it, what taxes and platforms take a share, and where regulation and influence sit. Use the output to map pricing, resellers, direct licenses, API tools, tokens, compute, and subscriptions. Run the same ledger again and again, because the scope, price range, and list of players can change as you validate them.
Pressure Testing Claims 31:00
He pushes the model to argue back and to act in a contentious, devil’s-advocate mode. He wants it to find holes, not agree with him. That lets him revise claims by evidence. A 5 to 10 billion figure drops to 1 billion. Another claim fails because the sources do not support it. He keeps only sourced points and throws out anything untagged or unsupported.
Independent Checks And Outreach 36:00
He then runs an independent checker across the project and gets a count of what can be verified, what is partial, and what is out of date. After that, he designs interview lanes for the uncertain claims and asks for contradictors in each one. He also uses his own data, contacts, and inboxes to draft outreach and identify warm connections, but he sends every message himself and warns against automated blasting.
Tools and workflows 47:30
He uses ChatGPT, Claude, and Gemini to scope questions, draft, and get hard ideas explained in plain words. For deeper work, he turns on research mode and higher thinking. He records calls with permission through Read.ai or similar tools, keeps notes in WhatsApp and email, and cross-checks them. The useful part is not the chat window. It is the standing tasks, daily summaries, trend scans, and alerts that keep each investigation moving until it needs his review.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.

