Which tool to choose
It works with any AI assistant: asking for multiple versions is an instruction, not a feature. For creative tasks (a title, a text, an idea) all the versions have value and you choose. For tasks with a correct answer (a calculation, a classification) it's better to ask for multiple independent reasonings and see which result recurs: the attempts that converge are more reliable than a single one. The choice of tool matters little; what matters is knowing which of the two cases you have in front of you.
How to do it
A single answer is a single roll of the dice: sometimes it comes out good, sometimes not, and you have no way of telling. Multiple versions give you a term of comparison, and for questions with an exact answer the version that recurs is almost always the right one.
- Ask for a precise number of different versions, not "a few alternatives": "give me four titles with different angles."
- For creative tasks, ask that the versions be genuinely different from one another, not variations of the same one.
- For tasks with a right answer, ask the AI to reason independently several times and then say which result recurs.
- Have it compare the versions and justify which one it recommends and why.
- You choose the final one: the AI's justification is an opinion, not a verdict.
The operational syntax, for a task with a correct answer:
Solve this problem three times, reasoning from scratch each time independently, without looking at the previous attempts. At the end tell me which result came up most often and is therefore the most reliable answer.
Problem: [text]
For creative tasks, instead, ask for the versions and then: "compare these four and tell me which works best for an audience of [...], explaining why." Read the justification, but trust your own judgment about the final destination.
A concrete example
Lorenzo has to choose a product name and the AI, on the first request, proposes one that doesn't convince him. Instead of relaunching at random, he asks for six names with different styles: one descriptive, one evocative, one short, one ironic. Then: "which two work best for a young audience and why." The AI argues for two, Lorenzo picks one and adds a variation of his own. From a single request he got six ideas and a line of reasoning, instead of a single name to take or leave.
When it does NOT work (and how to fix it)
If the versions are all the same as each other
The AI understood "alternatives" as "rephrasings." Impose the diversity explicitly: "each version must start from a different idea, not say the same thing in other words." Also indicate the dimensions to vary along: tone, length, angle.
If for a calculation different results come out on every attempt
This is exactly the case where repeating helps: ask for multiple independent reasonings and keep the result that recurs. If it doesn't converge even so, the problem is probably ill-posed or ambiguous: rephrase the question with more data, because the AI's uncertainty is reflecting an uncertainty in the question.
If you trust the version the AI says is the best and you're disappointed
The AI's justification holds for the measurable aspects, not for your taste or your context, which it doesn't fully know. Use its comparison to discard the weak versions, but the final choice is still yours: you're the only one who knows who it's really meant for.
A tip from someone who actually uses it
Always distinguish the two cases before asking. If the question has a right answer, multiple reasonings that converge give you reliability: use the repetition as a check. If the question is creative, multiple versions give you choice: use the variety as material. Confusing the two — asking for creativity where you need a right answer, or convergence where you need variety — is what wastes the attempts.
Frequently asked questions
Does asking for multiple versions consume more than usual?
Yes, an answer with four versions is longer than one with a single version, so it weighs more on your limit or your cost. But it's often worth it anyway: a single request with four versions costs less than four separate requests made by relaunching because the first one didn't work.
How does the version "that recurs the most" come to be the right one?
For questions with a correct answer, the AI's errors tend to be random and different on each attempt, while the right answer tends to recur. If across three independent reasonings two arrive at the same number, that number is more probable than the third, different one. It's not a guarantee, it's a filter that raises the odds.
Isn't it enough to redo the question until an answer I like comes out?
It's the slowest and most deceptive way. By relaunching at random you choose the version that confirms you, not the best one, and you lose the comparison. Asking for the versions together puts the alternatives in front of you all at once: you choose by seeing, not by hoping the next attempt is the good one.