Let AI interview you first to get dramatically better research

Executive overview

Most people try to craft better prompts themselves, but the bottleneck is rarely the prompt — it is the unstructured intent behind it. This video introduces a two-step method: use a fast reasoning model (O3 Mini High) to interview you one question at a time, then feed the resulting context-rich prompt into a deep research agent. The approach exploits the fact that reasoning models need the "what," not the "how," and that context is the single highest-value input you can give them. Comparing outputs across ChatGPT, Grok, and Perplexity shows meaningful quality differences that depend on the research topic.

Letting AI extract your intent before you research produces longer, more accurate, and more relevant reports than any manually crafted prompt.

The shift from prompt engineering to AI-prompted humans

• Current state: humans learn prompt engineering to get more from AI.
• Future direction: AI proactively interrogates users to surface implicit goals and constraints.
• This spectrum shift is already happening with agentic tools like deep research.
• The tactic in this video deliberately pulls that future forward today.
• Benefit is not just better output — it also helps users clarify what they actually want.

How the interview-then-research workflow works

• Start with a lightweight template prompt: name the topic, list known constraints, then instruct the model to ask one question at a time.
• Each answer must inform the next question — the model builds cumulative context, not a fixed checklist.
• O3 Mini High is chosen for speed; it handles multiple interview rounds without noticeable lag.
• At the end of the interview, the model writes a long, context-dense prompt automatically.
• That prompt is copied and pasted into a deep research tool (O1, Grok, Perplexity, etc.).
• Example used: helping a friend decide where to move, factoring in cost of living and crime data.

Prompting reasoning models vs. generative models

• Reasoning models (O1, O3, DeepSeek R1, Claude 3.7 Thinking, Grok 3) have chain-of-thought baked in.
• Generative models (GPT-4o, Claude 3.5, Gemini series) needed explicit chain-of-thought added by the user.
• Key rule for reasoning models: provide the "what," not the "how" — let the model decide the method.
• Prompt structure: goal (1–2 sentences) → optional response format → optional warnings → context.
• Context is the most important element; the interview method exists specifically to maximise it.

Choosing the right deep research tool

• ChatGPT deep research: longest reports and most consistent citations — best general-purpose choice.
• Grok: better for cutting-edge tech topics because it indexes X (Twitter), where AI discussion is concentrated.
• Perplexity: free or cheap, effectively unlimited queries — good for high-volume or exploratory use.
• ChatGPT Plus limits users to roughly 10 deep research calls per month; Pro unlocks ~50–100.
• Run the same prompt through multiple tools on important questions — outputs differ meaningfully.
• Length of output does not equal quality; citation accuracy is a stronger proxy for reliability.