Decision-making and tradeoff analysis

The task

You are weighing a real decision: a technical approach, a vendor choice, a hiring plan, or a strategic tradeoff. You want the model to help you see the consequences, not produce a generic pros-and-cons list.

Why the old approach is outdated

A common older prompt pattern forced a fixed analysis framework: list pros, list cons, score each option from 1 to 10, calculate a weighted total, and choose the highest number.

That can be useful when you already trust the criteria. It is a poor default when the important part is deciding which factors matter. Current reasoning models can compare options without being handed a rigid scoring sheet, and OpenAI’s reasoning guidance recommends concise answers with delimiters and constraints rather than asking for hidden chain-of-thought. For the broader concept, see chain of thought and reasoning models.

The current approach

State the decision, the real constraints, and how much risk you can tolerate. Ask for a recommendation plus the assumptions that would change it.

We're choosing between migrating our database to Postgres or staying on our current
managed MySQL setup.

Context:
- Team is 4 engineers.
- Nobody has deep Postgres operational experience.
- The driver is better JSON querying for a feature now in development.
- Migration budget is roughly 2 sprints.

Give me a recommendation for this situation specifically. Include:
- the strongest reason for the recommendation
- the biggest risk
- what assumption would change the answer
- the smallest test we could run before committing

The useful output is a decision record, not a transcript of the model’s private reasoning.

For higher-stakes decisions

Ask the model to separate known facts from assumptions and to name reversibility.

Before recommending an option, separate facts from assumptions. For each option,
state whether it is easy to reverse, expensive to reverse, or effectively one-way.

End with the decision you would make today and the evidence that would make you
revisit it.

If the model gives a confident recommendation without naming what could falsify it, the answer is not ready to use.

What to avoid

• Asking for “pros and cons” without constraints. It produces a general comparison, not a decision.
• Treating numeric scores as objective when the criteria are still uncertain.
• Asking the model to reveal chain-of-thought. Ask for assumptions, evidence, tradeoffs, and the final recommendation instead.
• Letting the model decide your risk tolerance. State whether speed, cost, reversibility, accuracy, or trust matters most.