Multi-Turn Prompting in ChatGPT for Iterative Work

Multi-turn prompting means treating the conversation as the tool: you give ChatGPT a task, read what comes back, and steer with follow-up messages until the output is right. The skill is in the follow-ups. Most marketers I work with put all their effort into the opening prompt, get a mediocre draft, and either accept it or start over in a fresh chat with a longer prompt. Both moves waste the thing the conversation already bought them, which is a model that now has the draft, the task and the context in front of it and can act on precise feedback about all three.

I think of it as the counterpart to prompt chaining. A chain is a pipeline you design in advance: angles, outline, draft, polish, each stage with one job. Multi-turn refinement is what you do when the work is exploratory and you cannot script the stages, because you will only know what is wrong with the output when you see it. Chains are for volume. Turns are for the one piece that has to be right. This post is the loop I run: why a follow-up beats a rewritten prompt, the five rules that keep a refinement loop converging instead of wandering, two worked examples, and the signs that a conversation is spent.

Multi-turn prompting versus prompt chaining

The two techniques get conflated because both involve a sequence of prompts, and the difference decides which one a task needs. In a chain, the sequence is planned before you start and each prompt has a defined input and output. In multi-turn work, the next prompt depends on what the last response got wrong, and you discover the sequence as you go. My shorthand: a chain is a recipe, a refinement loop is a conversation with a cook.

That difference tells you where each fits. When I generate twenty product descriptions, I want a chain with review gates, because the failure modes repeat and the fixes can be built into the stages. When I write one positioning statement, a homepage hero or a difficult apology email, I want a loop, because the first three versions exist to teach me what I actually want. Nobody can spec that in advance, and pretending to spec it is how you end up with a 400-word opening prompt that still misses.

Why a follow-up beats a rewritten prompt

A follow-up message carries information a fresh prompt cannot: the failed attempt itself. When you tell the model “version two, but the second paragraph drops the discount framing and leads with the deadline,” you are giving feedback anchored to a concrete artifact you both can see. Rewriting the opening prompt from scratch throws that anchor away and gambles that your new description will land better than your old one, when the whole reason you are iterating is that descriptions keep missing.

This is the same insight that makes few-shot prompting work, arrived at from the other side. Few-shot says examples beat descriptions for specifying what you want. In a refinement loop, the model’s own last output becomes the example, and your follow-up marks it up. “Keep the first line exactly as is, cut the third sentence, the tone of the closing is right but it runs long” is a marked-up example. I can rarely describe the email I want in advance, but I can always mark up the email in front of me, and so can every marketer I have taught this to. Editing judgment is the one skill the job already trained.

The five rules that make the loop converge

A refinement loop either converges toward something shippable or wanders in circles while the drafts get different rather than better. I run mine on five rules, and they exist because I have watched loops fail without them.

React to specifics, never restate the brief. The worst follow-up message is the original prompt paraphrased with “no, more punchy” attached. If the draft missed, quote the line that missed and say what is wrong with it. My test for a follow-up: could the model act on this message without guessing? “Make it better” fails the test. “The hook buries the offer until sentence three, lead with it” passes.

Pin what is already right. Models revise more than you asked. Tell ChatGPT to fix the closing and it will happily rewrite the opening you loved, because nothing told it the opening was settled. Every follow-up I send names what is frozen: “keep the subject line and the first paragraph exactly as they are.” Without pins, a loop can lose ground on turn four that it gained on turn two, and you will not notice until you compare versions.

One change per turn, mostly. When a follow-up asks for six fixes, you cannot tell which instruction caused which change, and the model has to weigh instructions against each other. I batch small mechanical fixes into one turn, but a judgment call - a tone shift, a restructure, a new angle - gets a turn to itself so I can evaluate it alone. Slower per turn, faster to done.

Ask for options at decision points, edits everywhere else. When I know something is wrong but I am unsure of the fix, I ask for three versions of just that piece: “give me three alternative opening lines, one leading with the deadline, one with the outcome, one with a question.” Options are for exploring a fork. Once the direction is chosen, I go back to single edits. Asking for three full drafts every turn is how loops sprawl.

Make the model surface what it is missing. Early in a loop, usually turn one or two, I ask: “what information do you not have that would change this draft?” The answer is regularly the fastest fix in the whole session, because a wrong assumption gets corrected once instead of leaking into every subsequent version. The model guessing your price point wrong will quietly poison ten turns of otherwise good feedback.

A worked loop: rescuing an email draft

Here is the shape of a real loop, compressed. The task is a win-back email for a lapsed customer segment, and it is hypothetical in its particulars, but the turns are exactly how I work. Turn one is a normal prompt with role, context and constraints, the anatomy every prompt in the prompt library shares. The draft that comes back is competent and dead: generic empathy opener, discount in paragraph two, “we miss you” energy throughout.

Turn 2:
The structure works but the opener is the same "we noticed you've
been away" line every brand sends. Cut it entirely. Open with the
concrete thing they lose access to next month. Keep everything
from "Your data is still here" onward unchanged.

Turn 3:
Better. Now the middle: "valued customer" appears twice, and the
discount is framed as a favor we're doing them. Reframe the
discount as "your old rate, held until the 15th." One version.

Turn 4:
What information do you not have that would make this email more
specific? List questions only, don't redraft yet.

Turn 5:
Good questions. Answers: they used the reporting feature most,
average tenure was 14 months, cancellation reason was mostly
"cost" not "switched." Now revise the middle paragraph only,
using the reporting detail. Subject line and opener are frozen.

Five turns, maybe six minutes, and the email is 90% shippable. Notice what never happened: I never rewrote the brief, never started over and never sent a follow-up the model would have to interpret. Each turn either marked up a specific passage, froze the settled parts or pulled missing facts into the loop. The question turn in the middle looks like a detour and is usually the highest-value message in the session.

A worked loop: tightening a positioning statement

Positioning work shows the option-generating side of the loop, because the problem early on is direction rather than polish. The pattern I run, again with hypothetical particulars:

Turn 1:
[Role, product facts, competitor list, the audience and the
change we want in how they describe us. Ask for one positioning
statement, 25 words max.]

Turn 2:
Too broad. "Powerful analytics for growing teams" could be any of
the five competitors I listed. Give me five versions, each one
built around a different fact from the context: the 10-minute
setup, the flat pricing, the no-SQL requirement, the audit trail,
the Slack-native alerts.

Turn 3:
Versions 2 and 5 are the only ones a competitor couldn't claim.
Merge them: flat pricing as the lead, Slack-native as the proof.
Two versions of the merge, one plain, one with more edge.

Turn 4:
The plain one. Now stress-test it: argue against this positioning
as our biggest competitor's CMO would. Three attacks, one line each.

Turn 5:
Attack 2 lands. Revise the statement so it survives it without
getting longer.

The stress-test turn is a role flip inside the same conversation, and it is something a pre-planned chain does badly, because I only knew the statement was worth attacking once turns two and three produced something defensible. That is the signature of multi-turn work: the best moves are reactions. I keep a handful of these reaction moves ready - attack it, compress it to half length, explain the reasoning behind the wording, say what you would test first - and deploy them when the draft in front of me earns one.

When to leave the conversation

Loops have a useful lifespan and then they decay, and knowing the exit signs matters as much as running the turns well. The practical problem with a very long session is that your early instructions and pinned decisions end up buried in the middle of a growing context, and the middle is exactly where models use information worst. The research on this is blunt: the “Lost in the Middle” study found that model performance is often highest when the relevant information sits at the beginning or end of the context and degrades significantly when it sits in the middle. In conversation terms, the rule you set on turn three is mid-context by turn twenty, and the model honoring it becomes a coin you should stop trusting.

My exit signs, from experience rather than the paper: the model reintroduces a phrase I banned ten turns ago, an old draft’s structure bleeds into a new variant, or a frozen section changes without being asked. Any one of these means the conversation’s working memory is failing, and more turns will fight the decay instead of improving the work.

The exit is cheap. I ask the conversation for its own handoff: “summarize the current approved draft, every frozen decision, and every rule I have given you, as a single brief I can paste into a new chat.” Then I start clean with that brief on top. Anything that should survive every conversation - product facts, claims rules, voice - does not belong in loop turns at all. It belongs in the system prompt layer, which rides underneath every new chat and never gets lost in the middle of anything.

FAQ

What is multi-turn prompting, in one sentence? Refining a model’s output through a sequence of follow-up messages in one conversation, where each message reacts to the last response, instead of trying to specify the perfect result in a single prompt.

How is it different from prompt chaining? A chain is a planned pipeline of separate prompts with defined stages, built for repeatable volume work. A multi-turn loop is unscripted refinement inside one conversation, built for single pieces where you discover what you want by reacting to drafts. I use chains for the batch and loops for the piece that matters most.

How many turns should a refinement loop take? My sessions mostly land between four and eight turns. Under three usually means I accepted something I should have pushed on. Past ten, I check for the decay signs above and usually restart with a handoff summary, because turn fourteen of an old conversation is worth less than turn two of a clean one.

Does this replace writing a good first prompt? No, and I still open with full role, context and constraints, because a strong first turn saves three corrective ones. The change is in where the effort peaks. I spend less time polishing the opening prompt to hypothetical perfection and more turns reacting to real output, since reacting is the cheaper and more reliable of the two.

This post is part of the prompt library. The planned-pipeline counterpart is prompt chaining, the technique for showing instead of describing is few-shot prompting, and the standing layer that survives every restart is system prompts. The wider practice starts at the complete guide to AI marketing.