The AI Question Method: Managing Frontier Models
Nate B Jonesgo watch the original →
the gist
Stop treating AI like a junior task-taker and start treating it like a senior partner by shifting from 'prompting' to 'questioning'—using intent-driven, multi-layered inquiries to guide reasoning across complex data sets.
The Shift from Prompting to Questioning
Prompt engineering as a standalone skill is now table stakes. With the advent of highly capable models like Opus 4.7 and OpenAI 5.5, the bottleneck is no longer the model's ability to execute, but the user's ability to define the work. The mental model must shift from giving tasks to a junior intern to partnering with a senior colleague. This requires moving away from rigid, task-based prompts toward a 'Question Method' that invites the AI to synthesize, analyze, and challenge your own assumptions.
Principle 1: The Flashlight Intent
Effective communication with an AI agent requires a 'flashlight' approach to intent. You must provide a clear, narrow center of focus (your thesis or core objective) while defining the edges of the problem space. By explicitly stating your perspective—even if it might be wrong—you give the AI a directional beam to work within. This prevents the model from wandering aimlessly and ensures it understands the boundaries of the investigation.
Principle 2: Synthesizing Complex Outcomes
Instead of relying on rigid evaluation scripts (evals) for every output, use your questions to force the AI to contend with the quality of the outcome. By asking layered, open-ended questions that require the AI to reconcile multiple competing variables (e.g., balancing customer emotion with technical feasibility in a PRFAQ), you leverage the model's reasoning capabilities to define what 'good' looks like. This collaborative wrestling with the problem is where the highest leverage is found.
Principle 3: Wrestling with Data and Opinion
When working with large context windows and multiple file types (transcripts, spreadsheets, PRDs), users often fail to force the AI to look across the entire data set. To avoid the model fixating on a single file, structure your questions to explicitly reference your data artifacts and your personal thesis. Ask the AI to synthesize a thesis that accounts for all provided inputs, specifically inviting it to agree or disagree with your assessment. This ensures the AI engages with the breadth of your context rather than just the most recent or prominent file.