Lesson 4 of 5 · 10 minutes · Module 2: Fix it and keep it

Fix an answer you don’t like

The first reply is only a first draft. In this lesson you’ll learn four ways to get a better one: say what’s wrong in the same chat, let the AI ask you questions first, give it permission to say it doesn’t know, and know when to start again.

You need

  • A free ChatGPT or Claude account
  • A reply you weren’t happy with (or the one from lesson 1)

A narrated walkthrough of this lesson is coming. The steps below cover everything in it.

  1. Step 1 of 6

    Say what’s wrong, in the same chat

    Reply in the same chat and say plainly what you want different. You don’t need to repeat the whole job. We told it the lesson 1 reply was a bit formal and asked for under 80 words. It came back shorter and more relaxed. It also noticed, without being asked, that the earlier version had promised something on someone else’s behalf, and took that line out.

    The prompt · works in ChatGPT, Claude or any AI chat tool
    That's a bit formal. Make it sound more like I'd talk to a neighbour, and keep it under 80 words.

    The reply we got, in the same chat as the specific prompt from lesson 1. Read our reply.

    Screenshot to come

    We haven’t captured this step yet. We’d rather show nothing than an old or made-up screen.

    CheckpointThe new reply fixes what you asked about.

  2. Step 2 of 6

    Change one thing at a time

    If several things are wrong, fix the most important first and check the result before you ask for more. OpenAI suggests breaking bigger jobs into smaller prompts and treating prompting as a conversation: you refine your request based on the answers you get.

    Screenshot captured 3 October 2026.
    1. Break big jobs into smaller prompts
    2. Treat it as a conversation and refine

    Step 2: Change one thing at a time

    OpenAI’s help page "How do I create a good prompt for an AI model?", showing the tips to right-size and prioritise requests and to work iteratively, treating prompting as a conversation.

    CheckpointYou’ve fixed the biggest problem before moving on to the next.

  3. Step 3 of 6

    Let it ask you questions first

    When you’re not sure what details it needs, ask it to ask you. We wanted a short “About me” section for the bakery’s website and asked for up to five questions before it wrote anything. It asked who’s behind the bakery, what makes it different, where it is, what tone to use and how long it should be. Fair questions, all of them.

    The prompt · works in ChatGPT, Claude or any AI chat tool
    I want to write a short "About me" section for my bakery's website. Before you write anything, ask me up to five questions you need answered. Ask them all at once, numbered.

    The five questions we got on 2026-10-03. Read our reply.

    Screenshot to come

    We haven’t captured this step yet. We’d rather show nothing than an old or made-up screen.

    CheckpointYou have a numbered list of questions from the AI.

  4. Step 4 of 6

    Answer them, then read the result

    Answer each question by number, in your own words. They don’t need to be polished. Ours came back at about 60 words, built from our answers. It left a gap for the email address, because we hadn’t given one, and said so.

    The prompt · works in ChatGPT, Claude or any AI chat tool
    1. I'm Jo. I started baking sourdough at home about five years ago, and began taking orders from friends and neighbours. 2. Slow-risen sourdough loaves and celebration cakes, made to order. Everything is baked in my own kitchen. 3. A home bakery in a small town. Collection only, on Friday and Saturday, 8am to 1pm. 4. Warm and personal, written as "I". 5. About 60 words, ending with an invitation to order by email.

    The “About me” section we got back. Read our reply.

    Screenshot to come

    We haven’t captured this step yet. We’d rather show nothing than an old or made-up screen.

    CheckpointYou have a draft built only from your answers.

  5. Step 5 of 6

    Tell it that it can say "I don’t know"

    AI tools tend to give an answer even when they’re unsure. Anthropic’s advice is to give Claude permission to admit uncertainty, which it says can greatly reduce false information. Add a line like “If something isn’t clear, tell me rather than guessing.” You’ll use this line in your prompt kit.

    Screenshot captured 3 October 2026.
    1. Give it permission to say "I don't know"

    Step 5: Tell it that it can say "I don’t know"

    Anthropic’s guide "Reduce hallucinations", at the first strategy: allow Claude to say "I don’t know" by explicitly giving it permission to admit uncertainty.

    CheckpointYou have a line you can add to any prompt where accuracy matters.

  6. Step 6 of 6

    Know when to start again

    If a chat has gone round in circles, stop. Start a new chat and write a better first prompt, using the four parts from lesson 2 and what you’ve learned from the replies so far. It can feel like giving up. It isn’t.

    CheckpointYou know when a fresh start beats another follow-up.

Common problems

When you have finished

Want to keep your place? Get a sign-in link by email and we’ll save your progress. You can keep reading either way. Sign in to save your progress.

Steps 2 and 5 are the makers’ public guidance pages, captured by tools/capture/prompting-for-beginners.mjs. Steps 1, 3 and 4 happen in a signed-in chat, so they have no screenshot yet; the prompts and replies are in outputs/.

Sources, checked 3 October 2026