By Joe Nalven Claude Gemini

This essay continues the exploration of using AI as a knowledge tool, covering a new prompt approach (thesis and inquiry for non-linear problems) and illustrated by applying it to the US-Iran conflict as of May 24, 2026.)

The Vending Machine Problem

For most of the short history of practical AI use, people have approached language models the way they approach a vending machine: insert a coin, press a button, receive a product. The prompt was a transaction. The AI was a dispenser. This was not entirely wrong — early models rewarded precise, bounded instructions. Give a vague command and you received a vague output. The discipline of “prompt engineering” grew up around this transactional logic, teaching users to specify context, constrain variables, define format, and anticipate failure modes before hitting enter.

That discipline still has its place. But something significant has shifted in 2025 and into 2026. Frontier models (the current generation of large language models operating at the highest capability tiers) are no longer well-served by the vending machine frame. They are capable of holding complexity, reconciling tension between competing goals, stress-testing assumptions, and surfacing the blind spots in a user’s own reasoning. They can do this, that is, if the person on the other end asks them to. Most people still don’t.

Nate Jones, whose prompting philosophy has attracted serious attention among knowledge workers and executives, has put it plainly: what used to be best practice is now table stakes. The mechanical discipline of structuring a prompt for autonomous execution (what Jones calls the “Four-Skill Framework”) was designed for a world where the model needed every variable pre-specified because it couldn’t handle ambiguity. That world is receding. The new challenge is not how to give a perfect order, but how to think out loud with a partner who is better at finding the flaws in your logic than you are.

Two Paradigms: The Technician and the Senior Partner

The distinction between the old approach and the new one is not merely stylistic. It reflects a fundamentally different theory of what AI is for.

Under the earlier framework, the human acts as an architect. Before the conversation begins, you have already defined the deliverable, constrained the scope, specified the format, and anticipated edge cases. You hand this blueprint to the model and it executes. The quality of output depends almost entirely on the quality of your pre-specification. If your blueprint is wrong, the model faithfully builds the wrong thing. It will do so with great competence and complete obedience.

Under what Jones calls the Thesis and Inquiry method or the AI Question Method, the human acts as a debater. You arrive with a working theory, not a finished specification. You share your hunch, name the tensions you see, and ask the model to push back. You invite it to find the counter-evidence in the data, to identify variables you haven’t accounted for, to surface the structural flaw in your reasoning before you’ve committed to it. The model is not executing your instructions; it is challenging your premises.

The practical difference shows up in where intelligence enters the conversation. In the command model, you bake your intelligence into the prompt before you hit enter. In the inquiry model, the intelligence emerges from the exchange. This matters enormously for complex, non-linear problems: the kind of problem that executives, analysts, journalists, and policymakers actually face — where the full specification doesn’t exist at the start and the problem itself reshapes as you investigate it.

Jones describes this........

© The Times of Israel (Blogs)