Ask for Three Versions, Then Choose

The president of a neighborhood association has to come up with the subject line for the email announcing the September flea market. She asks the AI for a suggestion and gets: "Join us for the big neighborhood flea market!" It's fine. She doesn't know whether it's good, since she has nothing to compare it to. She asks again, gets an almost identical variant, and ends up keeping the first one out of tiredness.

Deciding on a single answer amounts to accepting or refusing, rarely to choosing. With three truly different versions, you compare, and the comparison brings out what you want.

Ask for versions, not variants

If you only write "give me three suggestions," you often get three sentences that look alike: same words, same construction, a synonym here and there. The model tends toward the most probable wording and repeats it with small nuances.

The remedy is to say how the versions should differ, that is, to set angles.

I need to write the subject line of an email announcing our neighborhood flea market, on Sunday [DATE], at [PLACE]. The recipients are the residents on the association's mailing list.

Suggest three email subject lines, each with a different angle:
1. Informative: the date and place first.
2. Practical: what the reader needs to do (book a stall, come to browse).
3. Curious: a question or a detail that makes people want to open the email.

Each subject line is 60 characters maximum. The facts stay accurate from one version to the next: don't make up any times or activities.
Present the result in a table: angle | subject line | what this subject line bets on in the reader.

You choose the three angles, not the machine. For longer text, you can vary other axes: tone (sober, warm, direct), starting point (the benefit, the problem, an anecdote), length, or target reader.

The last column of the table forces you to read each version as a bet on the reader.

Why three and not ten

Three versions are enough to feel the differences. At ten, you're back to near-duplicates, and above all you have to read and sort, which cancels the time savings. If you don't like any of the three, don't ask for ten more: say what's missing, for example "too serious, lighter, no exclamation mark," as in the iteration method.

Choose, or mix

Once you have the three versions in front of you, there are four possible outcomes.

• You pick the best one and keep it as is.
• You take the structure of one and a word from another. This is common and effective: the curious version has a good turn of phrase, the informative version has the right order.
• You ask for a merge: "Keep the order of version 1 and the wording of version 3 for the ending."
• You reject all three, and now you know why, which gives you the next instruction.

To choose, don't rely only on your taste. Read each version aloud, imagine the hurried reader who sees it in their inbox, or show the three to someone who doesn't know the project. Two minutes of that opinion is worth more than a long debate with the AI, which doesn't know your readers either.

When it isn't worth it

Pointless when the answer is a fact, a calculation, or an extraction: there's only one right answer, and three versions would give you three chances to be wrong. The value appears when you have to choose a tone, an angle, or a wording. Rerunning the same request in new conversations also gives variations, less controlled ones: see parameters like temperature.

To do today

Take the next short text you have to write (email subject line, ad headline, first sentence of a letter). Write three angles that match what you're hesitating to say, run the prompt, then choose. Keep the angle you set aside in a corner: it will serve you for the next message.