Prompt AI to Choose Between Two Options

If you paste two options into a chat and ask which is better, you will get an answer, and the answer is unreliable for a reason that has nothing to do with the options. The model will tend to favour whichever one you described in more detail, and position in the prompt matters too. You wrote 200 words about option A and 60 about option B, so A sounds considered and B sounds like an afterthought. You have not asked a question. You have shown your hand and received it back.

Here is how to ask properly, and how to check whether the answer survived.

Step 1: write the criteria before you write the options

This is the whole technique. Everything else is refinement.

Ask for the decision criteria first, in a separate turn, with the options not yet mentioned. You want the model reasoning about what matters for this class of decision, not working backwards from a conclusion.

I need to choose between two options for how my small team stores
customer records. Before I describe them, list the 5 or 6 criteria
that should decide this, in priority order, and say in one line why
each one matters. Do not ask me what the options are yet.

Now you have a rubric that was written without knowledge of what it will judge. Paste it back in the next turn along with the options, and ask for a score against each criterion. The model still has to justify itself, but it is justifying against a standard it set before it could see which way to lean.

Step 2: describe both options at the same length

Count the words. If one option gets more, cut it back or pad the other out, because the imbalance alone will tilt the result.

The cleanest way is a fixed template per option, same fields, same order:

Option A
- What it is: ...
- Cost: ...
- Who maintains it: ...
- Main risk: ...
- Effort to reverse: ...

Option B
- What it is: ...
- Cost: ...
- Who maintains it: ...
- Main risk: ...
- Effort to reverse: ...

Leave a field blank if you genuinely do not know it. A blank is information. Filling it with speculation is how you end up with a confident answer about facts you invented.

Step 3: run the swap test

This is the step people skip, and it is the one that tells you whether to trust anything.

Run the same prompt twice. In the second run, swap the order so B appears first, and swap the labels so the option that was A is now called B. Same content, different position and name.

• Same winner both times: the recommendation is tracking the content. Trust it as far as you would trust any opinion.

• The first-listed option wins both times: the recommendation is tracking position. Throw it out, and go back to the criteria step.

• Different winners, both plausible: the options are genuinely close. That is a useful result. Stop looking for an answer from the model and pick on something it cannot weigh, like which one your team will actually maintain.

The swap test takes two minutes and it is the only part of this that produces evidence rather than text.

Step 4: force a decision rule, then apply it

Models are reluctant to commit. Ask which is better and you get a balanced survey ending in "it depends on your priorities", which is true, useless, and costs you a turn.

Fix that by asking for the rule rather than the verdict:

Give me a single decision rule in this form:
"Choose A if ___ is true. Otherwise choose B."
The condition must be something I can check today with information
I already have. Then tell me which branch my situation falls into,
and name the one fact that would flip it.

That last clause is the valuable one. A recommendation with a named flip condition is auditable. In six months when the situation changes you can check the condition instead of redoing the analysis.

What this will not fix

Two honest limits.

The model cannot know your context unless you tell it, and the things that usually decide real choices are unwritten: who on the team has done this before, what your last migration cost you in goodwill, which vendor your biggest customer already trusts. Those belong in the criteria you supply, not in the criteria the model generates.

And if you have already decided, this will not save you. A model is very good at building a case for a conclusion it can sense you hold. If you want the other side argued properly, ask for it directly, which is a different technique covered in prompting AI to argue against your own idea. The companion problem, a model that agrees with whatever you just said, is handled in stopping AI from being too agreeable.

When to stop prompting and automate

If you are making the same two-way call hundreds of times a day, routing a ticket, flagging a transaction, deciding whether to escalate, you have left the territory of prompting. That is a classification job, and a decision model returns a calibrated probability instead of a paragraph you have to parse. Prompting is for the decision you make once.

FAQ

Does the order of options really change the answer?

Often enough to matter. That is why the swap test exists rather than a rule of thumb. Run it and you will know for your prompt rather than guessing from somebody else's.

Should I give the model a recommendation to react to?

Only after you have the criteria. Lead with your preference and you have anchored the whole thing. If you want a reaction to your view, ask for the strongest case against it as a separate task.

Is it better to compare two options or five?

Two, in separate rounds. Comparisons degrade as options multiply because the model starts summarising rather than weighing. Run a bracket if you have five.

What if both options score the same?

Then the model has told you something real: the choice does not turn on the criteria you listed. Pick on reversibility. The option that is cheaper to undo wins ties.

For the broader toolkit, see the prompt engineering guide, and for when you cannot phrase the question properly in the first place, prompting when you do not know the right words.