Follow-up prompts: turn an OK answer into a great one
The first answer is only a draft. A short, specific reply is often all it needs. Here are the follow-ups worth knowing, and when to start a fresh chat instead.
In 30 seconds
- Treat the first answer as a draft, then say exactly what to keep and what to change.
- Ask what it left out, and say what to keep when you fix one part.
- Going round in circles? Ask for a summary and start a new chat with it.
You ask an AI for a few kind words for a colleague’s leaving card, and get something that could have come from any card in the shop. It’s tempting to copy it anyway, or give up. There’s a better option: reply. A follow-up is just another Jargon busterPrompt: What you type or say to an AI to tell it what you want., and it’s often where the good answers come from.
Treat the first answer as a draft
The first reply is the AI’s best guess, based only on what you’ve told it. Rather than starting again, reply as you would to a helpful colleague. It can see the conversation so far, so a short reply is enough, as long as it’s specific.
| Instead of | Try |
|---|---|
| “Make it better.” | “Make it warmer and half the length. Keep the line about the biscuit tin.” |
| “That’s wrong.” | “The leaving do is on Friday, not Saturday. Fix that and keep the rest.” |
| “Try again.” | “Give me three versions: one funny, one heartfelt, one very short.” |
| “I don’t get it.” | “Explain that again in plain English, with an everyday example.” |
| “Is that everything?” | “What have you left out that someone in my position should know?” |
Six follow-ups worth knowing
- Shorter or simpler“Cut this to half the length.” “Explain it as if I’m 12.” Say what must stay, so nothing important goes.
- A different tone“Warmer and less formal.” “Polite but firm.” “Sound like a person, not a press release.”
- Add examples“Give me two everyday examples of that.” Examples make an idea easier to grasp, and easier to check.
- Zoom in on one part“Expand point three into a full paragraph.” “Just the costs, please, in a table.”
- Ask what’s missing“What have you left out that someone in my position should know?” It often brings up the catches and exceptions.
- Change one thing only“Keep everything the same, except the date.” This helps stop it rewriting the parts you liked.
Not sure what’s missing? Turn it round: “What else do you need to know to make this better? Ask me one question at a time.” There’s more on this in let AI ask the questions first.
That’s close. Keep [the parts you like]. Change [what’s wrong, and why]. Don’t change anything else, and list what you changed at the end.
Push back the right way
If something looks wrong, say what and why, and ask it to check: “That total doesn’t match the receipts I pasted. Go through them line by line.” Be wary of a bare “Are you sure?” AI tends to agree with you, a habit called Jargon busterSycophancy: When an AI tells you what you want to hear, agreeing with you instead of giving an honest view., and researchers have found that models sometimes swap a right answer for a wrong one when challenged.
So point it at something it can check, such as the document you shared, rather than at your doubt. And invite it to disagree: “If I’ve got this wrong, tell me.”
When to start a new chat
Follow-ups work while a chat is on track. Start afresh when you’re going round in circles, with the same mistake coming back however you phrase it, or when you’ve changed topic and old details keep creeping into new answers.
There’s research behind this. A 2025 study from Microsoft Research found that when an AI takes a wrong turn early in a conversation, it tends to get lost and not recover. A fresh chat, with everything it needs in the first message, can do better.
Sum up this conversation as one prompt I can paste into a new chat. Include what I want, the key details I’ve given you, anything I’ve said I don’t want, and the best version so far. Leave out ideas we’ve dropped.
Fewer gaps in your first prompt means fewer follow-ups: here’s how to write a good prompt. And however polished the final answer, check the facts, because AI can still get things wrong.
Check yourself
3 quick questions nothing is savedTools in this guide
Sources (4)
- LLMs Get Lost In Multi-Turn ConversationMicrosoft Research, May 2025
- Towards Understanding Sycophancy in Language ModelsAnthropic, October 2023
- sycophancy-eval: datasets from “Towards Understanding Sycophancy in Language Models”GitHub
- Use Claude’s chat search and memory to build on previous contextAnthropic (Claude Help Center)
Spotted a mistake? Tell us and an editor will check it.
