The Two-Pass Prompt: Separate Exploration From the Final Decision

Many weak AI results begin with one overloaded request: “Give me ideas, compare them, choose the best one, write the plan and make it sound confident.” The model is being asked to open the field and close it at the same time. Those are different cognitive jobs.

A better method is the two-pass prompt. Pass One is divergent: it expands possibilities, exposes assumptions and finds missing questions. Pass Two is convergent: it evaluates the surviving options against explicit criteria and produces a decision-ready recommendation.

Why the one-shot prompt becomes muddy

When exploration and selection are compressed into one instruction, the first plausible idea can quietly become the winner. The answer may sound decisive without showing whether alternatives were seriously considered. It also becomes difficult for the human to see where evidence ends and model preference begins.

The problem is not simply “bad prompting.” It is a workflow-design problem. OpenAI’s current prompting guidance recommends clear context, smaller focused tasks and iterative refinement for complex work. Its writing guidance similarly frames useful AI work as a sequence—plan, draft, revise and package—rather than a single magical command.

Pass One: explore before you commit

In the first pass, forbid premature conclusions. Ask the AI to widen the option space, identify uncertainties and show what information would change the answer.

Exploration prompt: Generate six materially different approaches to [goal]. For each, state the target user, mechanism, main advantage, hidden assumption, likely failure mode and evidence we would need before choosing it. Do not recommend a winner yet.

The words “materially different” matter. Without them, an AI may produce six cosmetic variations of one concept. Asking for assumptions and failure modes prevents the list from becoming an advertisement for itself.

The human checkpoint

Do not rush directly into Pass Two. The human should remove impossible options, correct faulty context and add constraints the AI could not know. This is also the moment to label facts, estimates and preferences separately.

• Which options solve the real problem rather than the visible symptom?
• Which assumptions can be tested cheaply?
• What legal, financial, emotional or reputational limits were missing?
• Which choice would be difficult to reverse?

Pass Two: decide against a declared scorecard

The second pass should not ask, “Which is best?” Best according to whom, for what outcome and under which constraints? Give the model a scorecard.

Decision prompt: Compare Options A, C and F using these weighted criteria: user value 30%, implementation effort 20%, evidence strength 20%, reversibility 15% and downside risk 15%. Show the score, the reasoning, the strongest objection and the information still missing. Recommend one next experiment—not a permanent commitment.

This changes the AI from an answer generator into a structured comparison partner. The final output becomes easier to audit because the recommendation is tied to visible criteria.

A simple example

Imagine a creator deciding how to monetise a new audience. Pass One could explore subscriptions, digital products, sponsorship, licensing, affiliate sales and consulting. Pass Two might compare the top three using speed to revenue, recurring value, brand fit, workload and dependence on third-party platforms. The result is not merely an idea; it is a decision with a testable logic.

The rule to remember

Use AI to enlarge the map before you ask it to choose the road. Exploration without decision creates noise. Decision without exploration creates false certainty. The two-pass prompt gives each job its own space—and keeps the human responsible for the final commitment.

Further reading: OpenAI prompt guidance and OpenAI Academy’s Plan–Draft–Revise–Package workflow.