The follow-ups that turn an average answer around

A baker asks the AI for a text for the back-to-school newsletter: new walnut bread, Sunday brioches are back. The text arrives, it's smooth, it begins "Dear valued customers, it is with great pleasure that we..." Nothing is wrong, nothing makes you want to read on. Many people stop there and conclude that AI writes badly. Others type "redo it, but better," which gives the same text with one more adjective.

Between the two, there are eight follow-ups that work well because they tell the AI what to do with its own answer. The complete method of correcting in rounds is described in our guide to iterations. Here it's a repertoire: you look at what bothers you, and you pick the matching follow-up.

Which follow-up for which flaw

The follow-ups, one by one

Hunt down the generic

This is the most profitable one with lukewarm texts. Instead of demanding originality, have it point out the interchangeable passages.

Reread your text and underline every sentence that could appear as is in the newsletter of any bakery. List them, then rewrite the text replacing them with details specific to us: [OUR REAL DETAILS]. Invent no detail.

The last sentence matters: without it, the AI replaces the generic with the falsely specific.

Ask for different versions

"Give me three versions" often produces three barely touched-up copies. Impose the gap.

Propose three versions of the opening, built on completely unrelated angles: one that begins with a smell, one that begins with a practical piece of information, one that begins with a question to the reader. One sentence each.

You then pick the angle, and develop only that one.

Have it play the skeptical reader

The AI is more useful as a critic when you give it a specific character.

You are a customer receiving this newsletter and skimming it in ten seconds. Tell me where you stop reading, and what would make you want to keep going.

Keep in mind that an AI tends to find your text rather good. If the critique is soft, rephrase: "list the three weakest points, even if the text seems good to you." The phenomenon is described in the article on AI bias.

Lock what works

When a sentence pleases you, protect it, otherwise it will vanish in the next rewrite.

Keep the sentence "[YOUR SENTENCE]" and the last paragraph word for word. Rewrite everything between the two, in 60 words, without adding information.

Cut with a priority

"Shorten" gives a text amputated at random. Say what must survive: "Cut it in half. If you have to choose, keep the hours and the new bread, drop the rest."

Ask for its assumptions

When the result surprises you, don't argue: ask what the AI assumed. "What did you assume about my readers and my tone? List your assumptions." You often find out that it imagined big-city customers or a press-release tone, and you fix the source of the problem.

Have it list what needs checking

After a draft that contains facts: "List every piece of information in the text that you did not get from my request." Everything that comes out of that list comes from the AI, not from you. To go further, checking an answer has its own method.

Get it to ask you questions

If you sense that material is missing, rather than guessing which: "Ask me three questions whose answers would make this text more precise." The technique is detailed in the article on the questions the AI can ask you.

The limits of follow-ups

A follow-up corrects what the conversation already contains. If the first answer started from a wrong assumption, every follow-up builds on it. After three follow-ups that go in circles, open a fresh conversation with a prompt rewritten from what you've learned.

Also watch out for buildup: after seven or eight exchanges, the AI can mix old instructions with new ones. If a flaw you fixed comes back, it's a sign you're better off starting from zero.

To do today

Take an AI answer you judged "average" this week and give it the first follow-up from the table, the one that hunts down the generic. If the next answer finally contains details that come from you, keep that follow-up sentence in your library of reusable templates.