How to Get a Better Answer from AI: 4 Follow-Up Lines (2026)

You are reading Chapter 4 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.

To get a better answer from an AI assistant, reply to the one you got instead of starting again. The first reply is a draft, not a verdict, and four short follow-up lines fix most of what comes back: cut it down, aim it at the real point, ask for versions, or push for what it left out.

This guide gives you those four lines, the repair loop that makes them work, and the point where another correction is worse than a fresh chat.

We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.

Most people use an assistant like a vending machine. One message in, one reply out, and a judgement: good or useless. That judgement is being made on a first draft written by someone who has read one paragraph about your work.

Send a second line and the assistant re-reads the whole thread from the top, this time knowing what you did not like about the first attempt. That is not nagging it. That is the way it is built to work, and it is the single biggest difference between people who get value out of these tools and people who do not.

Take the reply that disappointed you in Chapter 3 and add one line to it rather than starting from a blank chat.

Four failures cover most first replies, and each has a line that repairs it.

The most common problem by far. Openings that restate your question, closings that summarise what you just read, hedging in the middle. Say "half that length", or "cut the introduction and the summary", or "less formal, write it the way I would say it out loud".

It wrote about the thing you asked about and missed what mattered. Name the point: "You missed the point. The point is that X. Try again."

When you want to choose rather than accept: "Give me 3 versions, different angles, one line each, no explanation." One line each is what stops you reading three essays to compare them.

Nothing is wrong with it and nothing in it is useful. Ask: "What did you leave out that I would regret not knowing?"

In all four, name the defect rather than describing your disappointment, fix one thing per message, and point at the exact sentence you object to.

The repair loop is four steps, and the discipline is in the first one.

One, not a list of five. A message carrying five corrections gets you a reply that half-addresses all of them.

No preamble, no apology, no explanation of why you want it. "Too long" is a complete instruction.

Did the defect you named get fixed? Everything else is the next round's problem.

Two corrections that go nowhere mean the problem is in the request, not the draft. Edit your original message to rewind the conversation, or start again.

There is a difference between correcting an output and feeding your requirements in one at a time. Research presented at ICLR 2026 found that spreading requirements across several turns cut performance by 39% on average. Say the whole ask at once, then repair what comes back.

Threads go stale long before they hit any technical limit. Old material sits in the same window as the new question and competes with it.

Stay in the thread when it is the same job and you are still tracking the same work. Start a new chat when the subject changes, when it revives a version you already rejected, when you find yourself re-explaining things you have said, or when you want an opinion that has not been shaped by everything above it.

A study of 18 leading models found that answers get steadily less reliable as the input gets longer, and that irrelevant material actively interferes. Once you change subjects, the old subject is not context, it is interference. Open a new chat and hand over three lines: what you are making, who it is for, and what you have already ruled out.

Before: you send one message, read a flat answer, decide the tool is overrated, and close the tab. Or you keep typing into the same thread for an hour and the answers quietly get worse.

After: you treat the first reply as a draft, name one defect, and steer it in a line or two. You know the four repairs by heart, and you know the two signals that mean stop steering and start a new chat.

One thing carries into Chapter 5: a request you had to repair more than twice. Chapter 5 works on the opening message instead of the repair - context, task and format, the three things a clear ask contains. If you want to go deeper on the asking side, our guide to the 26 principles of prompt engineering is the long version, and the rest of the syllabus lives in the AI Central Library.

Reply to the answer you got rather than starting a new chat. Name the single biggest thing wrong with it in one plain line - too long, wrong point, only one option, too shallow - and send that. The assistant re-reads the whole thread and tries again with that correction in hand.

"Half that length." If the padding is structural, add "cut the introduction and the summary". If it reads stiffly, "less formal, write it the way I would say it out loud" does more than asking for a shorter version.

Two. If two rounds of correction produce no real progress, the problem is in your original request rather than in the draft. Edit that first message to rewind the conversation, or open a new chat and restate the whole thing.

One thread per topic. Stay put while you are on the same job; start a new chat when the subject changes, when old rejected versions keep coming back, or when you are re-explaining things you already said. Answers get less reliable as the input gets longer, and irrelevant earlier material interferes with the current question.

All at once. Research presented at ICLR 2026 found that spreading requirements across multiple turns reduced performance by an average of 39%. Correcting an output over several messages is fine; drip-feeding the requirements is what costs you.