--- id: iterating lang: en title: "Iterating" seoTitle: "Iterating: refine an AI answer" description: "How to read Claude's first answer and steer the next one. The small follow-up prompts that turn a fair reply into a really good one, with worked examples." part: prompting partTitle: "Prompting" partOrder: 3 order: 3 level: Intermediate durationMinutes: 6 outcome: "Improve a fair answer into a strong one through short, targeted follow-ups." summary: "You rarely get the best answer first. Read the reply, then steer the next one with one small change at a time: shorter, sharper, or a different tone." draft: false updated: "2026-07-26" furtherReading: - title: "Prompt engineering overview (Anthropic)" url: "https://docs.claude.com/en/docs/build-with-claude/prompt-engineering/overview" note: "includes tips on refining and steering answers." - title: "Anthropic's interactive prompt engineering tutorial" url: "https://github.com/anthropics/prompt-eng-interactive-tutorial" note: "practice troubleshooting prompts step by step." - title: "Prompt engineering guide (OpenAI)" url: "https://platform.openai.com/docs/guides/prompt-engineering" note: "more strategies for getting from a rough answer to a right one." keywords: - prompt iteration - refine AI answer - follow-up prompts - improve AI output - how to fix AI answers quiz: intro: "Four quick questions. Clear them all and the next lesson opens up." passScore: 4 questions: - q: "The first answer is only okay. What's the best next move?" options: - "Start a brand-new chat and hope for better" - "Tell it what to keep and what to change, in one specific follow-up" - "Ask it to 'do better'" - "Copy it into a different AI tool" answer: 1 explain: "Steer it. Name what works, name what doesn't, and it fixes the gap instead of guessing again." - q: "Which follow-up will actually move the answer?" options: - "That's not quite right." - "Make it good." - "Cut it to three sentences and drop the last paragraph." - "Hmm, try again." answer: 2 explain: "Specific and concrete. The AI can act on 'three sentences, drop the last paragraph.' It can't act on a vibe." - q: "When is it smarter to start fresh than to keep steering?" options: - "Never. Always keep the same chat going." - "When the chat has drifted far off track and each fix makes it worse" - "After exactly three follow-ups, every time" - "Only if the AI apologises" answer: 1 explain: "When you're fighting the thread more than the task, start clean with a better opening prompt." - q: "Why change one thing at a time instead of ten?" options: - "It's polite to the AI" - "So you can see what each change did and keep what works" - "It uses fewer words" - "There's no reason; change everything at once" answer: 1 explain: "One change at a time tells you what actually helped. Ten at once and you're back to guessing." pass: badge: "Perfect score" title: "You're steering now, not gambling." body: "Full marks. You can take a fair answer and coach it into a good one. Next: patterns for specific kinds of work." retry: "Check the marked answers and give it another go. You've nearly got it." --- > **Level:** Intermediate ยท **Reading time:** ~6 min > > **What you'll be able to do:** improve a fair answer into a strong one through short, targeted follow-ups. Here's a habit worth breaking: you send a prompt, the answer comes back a bit off, and you either sigh and use it anyway or wipe the slate and start over. Both waste the best thing the AI gives you, which is a first draft you can react to. **The first answer is rarely the finished one. It's a starting point.** The skill is reading it, spotting the one or two things that are off, and steering the next version. This is where a lot of the real quality lives. You already wrote a solid prompt in the [last lesson](/learn/prompting/how-to-write-a-great-prompt). Now you learn to *coach* the reply. ## Read the first answer for what to fix Before you type anything, actually read what came back and ask three quick questions. Is it the right length? Is the tone right? Is anything wrong or missing? Most first drafts are close on one, off on another. Maybe the facts are right but it reads like a brochure. Maybe the tone is spot on but it's twice as long as you need. **Naming the gap is half the work,** because once you can name it, you can ask for the fix. ## Steer with small, specific follow-ups You don't need to rewrite the prompt. You just tell the AI what to keep and what to change. The trick is being concrete. Vague steering gets you nowhere: ``` Make it better. ``` Better how? It'll shuffle a few words and hand you back much the same thing. Specific steering works: ``` Good, but two fixes. Cut it to three sentences, and lose the "we're excited to" opening. Get straight to the ask. ``` That, it can do. "Three sentences" is a number. "Lose the opening" is a clear target. The more precise the instruction, the closer the next version lands. Think of it as **pointing, not hoping.** ## Watch it work Say you asked for a note declining a meeting and got back something warm but far too long. Your first follow-up: ``` Right idea, too long. Half the length, and keep it friendly. ``` Shorter now, but it dropped the reason you're saying no. Second follow-up: ``` Better. Add one line: I'm out because I'm heads-down on the launch this week. ``` Now it's short, friendly, and honest. Two small nudges, thirty seconds, and a fair draft became one you'd actually send. Notice you changed *one thing* each time. That's on purpose.
One change per step lifts the draft A fair first draft rises to a ready-to-send note in two steps, cutting the length first, then adding the reason, one change at a time. First draft long, warm cut the length Shorter but missing the why add the reason Ready to send short, friendly, honest One change per step, so you can see what each nudge did.
Steering a draft across a few turns A chat thread. Claude's long first draft, then your follow-up to cut it in half, a tighter reply, your follow-up to add one line about the launch, and the final ready-to-send note. One change per turn. Draft 1: warm, but eight paragraphs long. Cut it to half. Keep it friendly. Draft 2: tighter, but a bit flat. Add a line: heads-down on the launch. Draft 3: short, friendly, honest. Ready to send ✓ You in teal (right), Claude in the light bubble (left): five turns, one change each.
## One change at a time It's tempting to fire off "shorter, warmer, add the numbers, fix the ending, and make it funny" all at once. Resist it. When you change five things together and the result is worse, you've no idea which change did the damage. **Change one thing, see what it did, keep it or undo it, then change the next.** It *feels* slower. It's faster, because you're learning what works instead of rolling the dice again. ## When to steer and when to start fresh Steering has a limit. Sometimes a chat drifts so far that every fix drags in the baggage of the wrong turns before it, and you spend more effort fighting the thread than doing the task. That's your cue to **start clean.** Open a new chat and write one strong opening prompt that folds in everything you've learned from the messy one. You're not throwing away the work: you're using what the failed attempt taught you. Knowing *when* to steer and *when* to restart is its own small skill, and it comes fast with practice. ## Try it Take a so-so answer you have lying around, or generate one on purpose with a lazy prompt. Improve it in exactly three follow-ups, changing one thing each time: first the length, then the tone, then one specific fix. Read the final version against the first. Three small nudges, one at a time, and watch how far a middling draft can travel. ## Key terms [Prompt](/glossary#prompt), [Iteration](/glossary#iteration), [Context](/glossary#context)