B5 · Generative AI and Conversational Assistants

Improving Results Through Follow-Up

Follow-up means continuing the conversation to improve an answer: asking for changes, adding detail you left out, correcting mistakes and requesting alternatives. Because the tool sees the whole conversation, each message can build on the last.

8 min Beginner Jump to the activity

What you need: Nothing. This page starts from the beginning, with no programming and no formulas.

Definition

The first answer is not quite right. Do you start again, or is there a better way?

Follow-up means continuing the conversation to improve an answer: asking for changes, adding detail you left out, correcting mistakes and requesting alternatives. Because the tool sees the whole conversation, each message can build on the last.

In simple words

When a tailor shows you a half-finished kurta, you do not throw it away and order another. You say "a little longer here, looser at the sleeve", and it is adjusted. Working with a chat assistant is the same. The first answer is a first fitting. A few clear follow-ups usually get you to something that fits.

Why it matters

Very few people write a perfect prompt the first time, and they do not need to. Treating the exchange as a conversation is faster than rewriting the prompt from scratch, and it produces better results. Follow-up is also how you find and fix the tool's mistakes.

Key terms

Follow-up
A later message that builds on an earlier answer.
Iteration
Improving something through repeated rounds of try, review and adjust.
Refine
To make an answer more precise or better suited to your need.
Conversation context
Everything said so far in the chat, which the tool uses when it replies.
Sycophancy
A tool's tendency to agree with the user, even when the user is wrong.

Explained in detail

A conversation, not a vending machine

A search engine works like a vending machine: you put in a few words and take what comes out. A chat assistant works differently, because it keeps the whole conversation in view. You can say "make it shorter" and it knows what "it" refers to. You can say "use the second idea" and it knows which idea came second. Many people use chat assistants as if they were search engines, typing one question and leaving. They miss most of what the tool can do.

Six useful kinds of follow-up

KindWhen to use itWhat to say
NarrowThe answer is too general"Focus only on the causes, not the effects."
ExpandOne part needs more"Tell me more about the third point, with an example."
ReshapeThe form is wrong"Turn this into a table." "Make it half as long."
Re-pitchIt is too hard or too childish"Explain it more simply." "Use a more formal tone."
CorrectSomething is wrong or missing"The date in the second paragraph is wrong. It was 1947, not 1948."
VaryYou want other options"Give me three different openings to choose from."

Ask the tool to examine its own answer

Some of the most useful follow-ups turn the tool on its own work. Ask: "What are the weakest points in this answer?" "What might a critic say?" "Which of these facts are you least sure about?" "What did you leave out?" The tool often identifies real problems, because spotting faults in a text is easier for it than avoiding them while writing. This does not replace your own checking, but it is a quick way to find where checking is needed most.

The trap of the agreeable assistant

Chat assistants are trained to be helpful and pleasant, and one side effect is that they tend to agree with you. If the tool gives a correct answer and you reply "I think that's wrong", it may apologise and change to an incorrect one. If you ask "Are you sure?" it may switch its answer simply because the question suggests doubt. Researchers call this tendency sycophancy. It means that a changed answer is not evidence that the first one was wrong.

The remedy is to ask for evidence, not agreement. Instead of "Are you sure?", ask "What is the source for that?" or "Show me the working step by step." Instead of telling it your opinion first, ask the question neutrally. Then check the evidence yourself.

When to start a new chat

  • The conversation has become very long, and the tool is forgetting or mixing up earlier parts.
  • The tool is stuck on a wrong idea and keeps returning to it.
  • You are moving to a different subject, and the old context would only confuse matters.
  • You want a fresh opinion that is not coloured by what was said before.

When you do start again, carry over what you learned. Write a new first prompt that includes the details the earlier conversation showed were needed.

How it works

  1. Read the answer properly before reacting. Decide what is good and what is not.
  2. Name the problem specifically: too long, too vague, wrong level, a wrong fact.
  3. Ask for one change at a time, so that you can see its effect.
  4. Keep what works: "Keep the first paragraph exactly as it is and rewrite the rest."
  5. Stop when it is good enough, and check the facts before you use it.

Try it yourself

Three turns to a better answer

A student wants help explaining why the seasons change, for a short talk.

Put the steps of refining an answer in order

What is a sensible sequence?

  1. 1Write a clear first prompt
  2. 2Check the important facts yourself
  3. 3Decide what exactly needs to change
  4. 4Ask for one specific change
  5. 5Read the whole answer

Common mistakes

  • Giving up after the first answer, or retyping the same prompt hoping for something different.
  • Saying "make it better" without saying what is wrong.
  • Taking a changed answer as proof that the tool has corrected itself. It may only be agreeing with you.

Check yourself

Question 1

1. Why can you say "make it shorter" to a chat assistant without repeating the text?

Question 2

2. Which follow-up is most useful?

Question 3

3. You tell the assistant its correct answer is wrong, and it apologises and changes it. What does this show?

Summary

  • Treat a chat assistant as a conversation: build on each answer.
  • Good follow-ups are specific and change one thing at a time.
  • Ask for evidence, not agreement, and start a new chat when the old one has lost its way.

Finished reading?

Sources: UNESCO: Guidance for generative AI in education and research · AI4K12: Five Big Ideas in AI