What it is
Iterative refinement means improving an AI's answer over several rounds instead of expecting it to be right the first time. You ask for something, read what comes back, tell the model what to change, and ask again. Each round moves the result a little closer to what you actually want. It's the same way you'd work with a new assistant, a freelance designer, or an editor: first draft, notes, second draft, more notes, done.
The word "iterative" just means repeated. "Refinement" means making something better in small steps. Put together, the idea is simple: the first answer is a draft, not a verdict.
This article is about the version where you are the one giving the feedback. There is also an automated version, where the model critiques and rewrites its own work in a loop. Researchers have studied that closely (the best-known paper, "Self-Refine", published in March 2023, even has "iterative refinement" in its title), and it's covered in self-critique. The human version matters more for most people, because it's what you do every time you reply to a chatbot with "Close, but…".
Why doesn't the first answer just come out right? Mostly because you haven't told the model everything yet, and often you don't know everything yet either. When you ask for "a description of my handmade candles", the model has to guess the length, the tone, who's buying, what makes your candles different, and what you'd hate to see. It fills those gaps with the most typical choices. The result is usually competent and generic. Reading it is often the moment you discover what you wanted: "no, not cosy, these are for people who hate cosy", "mention that they're vegan", "shorter". The first draft is partly a tool for finding out your own requirements.
That's why the skill in iterative refinement isn't typing "try again". It's turning your reaction into instructions the model can act on. "I don't like it" gives the model nothing to work with. "Cut the first sentence, replace 'luxurious' with something plainer, and add one line about the 60-hour burn time" gives it three clear jobs.
Refinement also has a shape. Early rounds fix the big things: the angle, the audience, the structure. Middle rounds fix the details: word choice, length, the order of points. Late rounds should be few and small. If you're still changing the angle in round seven, the problem isn't the draft. It's the brief, and a fresh start will probably serve you better than an eighth round.
Why it matters
Almost nobody gets the best result from one prompt. Even people who write excellent prompts usually go back and forth two or three times on anything that matters. Treating the first answer as final means accepting the model's guesses about all the things you didn't say.
It's where your judgement goes in. A model can produce a fluent paragraph in seconds. It can't know that your readers are retired teachers, that you promised never to use exclamation marks, or that the price is changing next month. Each round of feedback is a chance to put your knowledge, taste, and standards into the work. Without those rounds, the output reads like everyone else's, because it's built from the most common choices.
It keeps you checking. Every round forces you to read the draft properly. That matters because AI drafts can be wrong in confident, polished ways. People who accept the first answer tend to accept its mistakes too. The consultant study in the examples below found exactly that on a task the AI got wrong.
It's faster than starting over, up to a point. Fixing a nearly good draft with one precise note is quicker than rewriting a prompt from scratch. But the point matters. Long conversations have their own problems, and knowing when to stop refining and start fresh is part of the skill, covered in detail below.
It teaches you what to put in the prompt next time. The notes you give in rounds two and three are the details missing from round one. Write them down, and your next first draft starts where this one finished. Over time, repeated notes become part of a saved prompt template or a system prompt, and you need fewer rounds.
Real-world examples
Comparing drafts beats polishing one at a time (Stanford, 2010)
Long before chatbots, design researchers were studying how feedback rounds work. In a study published in December 2010 in ACM Transactions on Computer-Human Interaction, Steven Dow, Scott Klemmer and colleagues at Stanford asked 33 people to design a web banner ad for Ambidextrous, a student-run design magazine. Everyone made five practice versions and then a final ad, and everyone got the same kind of written critique on each practice version, from a list of about 50 statements written in advance by advertising and graphic design professionals.
The only difference was timing. The "serial" group got feedback after every single version, one at a time. The "parallel" group made three versions, got feedback on all three together, made two more, got feedback again, and then made their final ad.
The researchers then ran all 33 final ads as a real 15-day campaign on MySpace, the social network, and measured clicks. The parallel group's ads did better on every measure: click-through rate, time visitors spent on the magazine's site, and ratings from four magazine editors and three advertising professionals who didn't know which group made which ad. Their practice versions were also rated as more varied. They reported a bigger gain in confidence, and in interviews nearly half of the serial group reacted negatively to criticism of their work, while none of the parallel group did.
The lesson carries straight over to working with AI. If you refine a single draft round after round, you can get stuck polishing one idea. Asking for three different versions first, comparing them, and giving feedback on what works across them tends to lead somewhere better. Chatbots make this almost free: "Give me three versions that take different angles" costs one message.
What happened when consultants stopped questioning the draft (September 2023)
In September 2023, researchers from Harvard Business School, Wharton, MIT Sloan, Warwick Business School and Boston Consulting Group released a working paper called "Navigating the Jagged Technological Frontier". They ran an experiment with 758 BCG consultants, about 7 percent of the firm's individual-contributor consultants, some with access to GPT-4 and some without.
On 18 realistic consulting tasks that the AI was good at, such as generating product ideas and drafting marketing copy, consultants with AI finished 12.2 percent more tasks, worked 25.1 percent faster, and produced work rated more than 40 percent higher in quality. But the researchers also set one task they deliberately chose to be outside what the AI could do well: a business problem that needed careful reading of interview notes alongside spreadsheet data. Consultants without AI got it right about 84.5 percent of the time. The groups with AI got it right about 60 and 70 percent of the time, an average drop of 19 percentage points.
The paper's explanation is the part that matters here. The consultants who did worse with AI tended to adopt its output as it came and question it less. The AI's answer was persuasive and well written, so they stopped refining and started accepting. The authors also described two styles among people who used AI well. "Centaurs" divided the work, handing some parts to the AI and doing others themselves. "Cyborgs" worked with the AI continuously, going back and forth at the level of individual sentences and steps. Both styles involve reading every draft critically and pushing back. Neither involves taking the first answer on trust.
The study has limits: it was one firm, one AI model, and tasks designed by the researchers. But it's strong evidence that the habit of treating AI output as a draft to interrogate is not a nicety. On the tasks where the AI is quietly wrong, it's the whole difference.
Chatbots add a place to edit a draft in rounds (2024)
In 2024 the two best-known chatbot makers each added a feature built around refining a draft rather than chatting about it. In June 2024, alongside its Claude 3.5 Sonnet model, Anthropic introduced Artifacts: when you ask Claude for a document, some code, or a design, it appears in its own window beside the conversation, where you can see it, keep asking for changes, and build on it. In October 2024 OpenAI launched canvas for ChatGPT, a similar side panel for writing and coding projects. With canvas you could highlight one sentence or paragraph and ask for a change to just that part, or use shortcuts to adjust the length or reading level, ask for suggested edits, or give the piece a final polish.
What both features recognise is that a draft scrolling away in a long chat is awkward to refine. Pointing at the exact sentence you want changed is a much more specific kind of feedback than describing it. These products change often. At the time of writing, OpenAI has replaced canvas with editable writing blocks inside the chat itself, and both companies keep adjusting how drafts, versions, and edits work. Check what your own chatbot offers today. The idea underneath, a stable draft you can point at and improve piece by piece, is what matters.
Why long refinement threads go wrong (May 2025)
A brief note on the research, because the context engineering article covers it in full. In May 2025, Philippe Laban and colleagues at Microsoft Research and Salesforce Research tested leading models on tasks fed to them in pieces over several turns, the way people actually chat, instead of all at once. Across more than 200,000 simulated conversations, performance dropped by 39 percent on average. Models made early guesses, committed to them, and struggled to recover once they'd taken a wrong turn. For iterative refinement, this is the scientific case for a habit that experienced users already have: when a thread has gone off track, gather what you've learned into a clean brief and start a fresh chat.
An everyday scenario: the About page that got worse before it got better
This is an illustration, not a documented case. Priya runs a small dog-grooming business and asks a chatbot for the About page on her website. Round one is generic: "passionate team", "your furry friend deserves the best". She replies with three specific notes: it's just her, not a team; mention that she trained as a veterinary nurse; and ban the phrase "furry friend". Round two is much better. Round three fixes the length. In round four she asks for "a bit more warmth", and the model rewrites the whole page, including her favourite opening line, and slips "furry friend" back in.
She doesn't argue with round four. She scrolls up, copies round three into her notes as the best version so far, and asks for one change only: "Take version 3. Keep every sentence as it is except the last one, which should invite people to book a first visit. Change nothing else." Round five is the one she publishes.
Try it yourself
This exercise refines a book blurb (the short description on the back cover or a store page) in three planned rounds. If you don't have a book, use a product, a service, a course, or an event. You need any chatbot and about twenty minutes.
- Write down your non-negotiables before you start. Three to five facts the draft must get right (title, genre, main character, the core conflict, the length limit) and anything you never want to see. For example: "under 150 words; the heroine is 62, not young; no 'in a world where'; don't give away the ending".
- Round one: ask for three different drafts, not one. The Stanford study above suggests comparing options beats polishing the first one. Paste this, filling in the brackets: “
Write three different blurbs for my book. Make each one take a clearly different angle (for example: lead with the character, lead with the mystery, lead with the setting). Each must be under [length] words. Title: [title] Genre: [genre] Main character: [who they are, in one line] The problem they face: [one or two lines] What readers of this genre expect: [one line] Never include: [your banned phrases or spoilers]“ - Pick one and name what works. Don't just pick. Say why, because what you liked in the losing drafts is useful too. For example: "Draft 2 is closest. Keep its opening line. I like the humour in draft 3's last sentence."
- Run the feedback checklist. Before writing your round-two notes, check the chosen draft against these questions, and note the answer to each:
- Are all my non-negotiable facts right? (Check each one yourself.)
- Did it add anything I didn't give it, such as an invented detail, a quote, or an award?
- Is it within the length limit? (Count it yourself or use a word counter. Models are unreliable at counting.)
- Which one sentence is the weakest?
- Does it sound like the kind of book it is?
- Is there anything a reader would misunderstand?
- Round two: give specific, limited feedback. Turn your checklist answers into two or three concrete changes, and say what to keep. Use this pattern: “` Revise draft [number]. Make only these changes:
- [a specific change, e.g. "Replace 'a mysterious stranger' with 'her late husband's business partner'"]
- [a specific change]
- [optionally, borrow from another draft: "End with draft 3's last sentence"] Keep everything else word for word. After the blurb, list what you changed. “` Asking it to list what changed lets you spot any change you didn't ask for.
- Save the best version so far. Copy round two's blurb into a note, labelled "v2". From here on, you always have something to go back to.
- Round three: polish one thing. Pick the single weakest remaining point, and ask for that alone: “
Keep this blurb exactly as it is, except the [first / last / third] sentence. Give me four alternative versions of that sentence only, each under [number] words.“ Choose one of the four, or write your own using the best parts. - Decide whether to stop. Put v2 and your final version side by side. If the final is clearly better, you're done. If you can't tell, or it's worse, go back to v2. Three focused rounds are usually enough for a short piece. If it still isn't working, see the next step.
- If it went wrong, start fresh with a clean brief. Ask the chatbot to "Summarise everything I've told you I want in this blurb as a single brief, including the facts, the tone, and what to avoid." Check the summary, fix anything wrong, and paste it into a new chat. That brief is also the start of a reusable prompt for your next book.
Other feedback phrases worth keeping:
Don't rewrite it. Tell me the two weakest sentences and why.
This is close. The only problem is [X]. Fix that and leave the rest alone.
You changed [a sentence I liked]. Put back the original wording and keep your other changes.
Common mistakes and misunderstandings
Giving vague feedback. "Make it better", "more engaging", "try again" leave the model guessing again, so it changes things at random and often makes the whole piece longer and blander. Name the sentence, the word, or the quality, and say which way to move it. The Self-Refine researchers found the same thing for model-written feedback: on the task where they tested it, specific notes clearly beat generic ones. Your notes work the same way.
Changing everything at once. If you ask for a new tone, a shorter length, a different opening, and a stronger ending in one message, you won't know which change caused the new problems. One to three clear changes per round is plenty. It's the same one-change-at-a-time rule that evals use for testing prompts.
Not saying what to keep. Models tend to rewrite the whole piece when asked to change part of it, and good sentences are lost along the way. Say "keep everything else word for word", or point at the exact part to change. If your chatbot lets you highlight a section, use that.
Losing the best version. It's common for round five to be worse than round three, and in a long chat round three scrolls out of view. Copy any version you'd be happy with into a note before asking for more. Then going back costs you nothing.
Arguing with a thread that's gone wrong. When a model keeps making the same mistake after you've corrected it twice, more corrections rarely help. The mistake is now part of the conversation it's working from. Start a new chat with a clean summary of what you want, including the corrections, and you'll usually get it right in one go.
Replying when you should edit. Most chatbots let you edit a message you already sent and run it again from there. If the problem is something you left out of your original request (the audience, the length, a key fact), editing the original prompt is often better than adding a correction below. The model then works from one clear instruction rather than an instruction plus a patch. Replying in the thread is better when you're reacting to what it wrote. Check how your chatbot handles edits: some keep the old version so you can switch back, and some replace it.
Pushing back without a reason. "Are you sure?" invites the model to change a correct answer to please you, as the research described in self-critique shows. If you think something is wrong, say what and why, or give it the source to check against.
Not knowing when to stop. After a few rounds, the changes get smaller and more arbitrary, the style drifts away from what you liked, and fixes start to undo earlier fixes. That's the signal to stop, pick the best version you've saved, and finish it by hand. The last few percent of a piece of writing is often quicker to do yourself than to describe.
Treating a polished draft as a checked one. Every round makes the text smoother, and smooth text is easier to trust. None of that checks the facts. Before anything goes out, read it once purely for accuracy, the way the consultants in the BCG study who did well kept questioning the output. For anything that matters, keep a human in the loop, and that human is you.
Related elements
- Self-critique: the automated cousin, where the model reviews and revises its own work.
- Evals: testing a prompt against a set of cases, the way to refine a prompt you'll use again and again.
- Prompt: the starting instruction, which better refinement notes teach you to write.
- Prompt Templates: where your repeated feedback notes end up, so the next first draft starts closer.
- Prompt Chaining: splitting a job into planned steps with checks between them, for work too big for one draft.
- Context Engineering: why long, messy conversations degrade, and how a clean brief fixes it.
- Context Window: the limit on how much conversation the model can take in, which long refinement threads eat into.
- Writing Assistance: using AI on your own writing without losing your voice, where rounds of small edits matter most.
- Temperature: the setting behind why the same request gives different drafts, which is what makes asking for alternatives useful.
- Hallucination: the invented details that can survive, polished, through every round.
- Human in the Loop: the principle of a person making the final call, which is what each round of feedback is.