Stop prompting once: the iterative conversation method

There is a moment early in everyone’s AI journey when the model responds with something disappointing and they do one of two things: shrug and accept it, or open a new conversation and rewrite the prompt. If the response is close enough to diagnose, a better move is often to reply and steer the work.

What follows is a workflow for doing that deliberately. We will go through what to do when the first response is close, when it is off, and when it is wrong in a way that looks fine on the surface. By the end you will have a repeatable way to improve and check a draft without relying on clever wording.

Why most people stop too soon

The vending-machine model of AI, type in a prompt and get out an answer, misses what a chat interface can do. Follow-up turns add instructions and reactions to the working context. When the first response is a useful starting point, that context lets you identify what to keep and what to change.

This can help even after a strong first prompt because later turns include your reaction to the draft. Vendor guidance describes draft, review, and refine as a common prompt-chaining pattern, while also recommending explicit criteria and separate steps when the intermediate review must be inspected or logged (Anthropic’s prompting best practices).

It can feel slightly impolite to keep pushing back, or inefficient because you think the first prompt should have been perfect. You are editing a generated draft, not negotiating with a person. Continue while each turn makes a specific, reviewable improvement; restart when the accumulated context is actively getting in the way.

The four moves

This workflow uses four kinds of follow-up before the final verification pass.

1. Critique: tell the model what is wrong, specifically.
2. Narrow: ask for one piece of the response, expanded.
3. Pivot: change the angle or the audience.
4. Stress-test: ask for assumptions, evidence, failure cases, or a credible counterargument.

We will go through each.

Treat these four moves as a reusable workflow, not a bag of clever follow-up prompts. For recurring work, save the sequence as a template your team can reuse.

Move 1: Critique

Use critique when the first response is mostly good but specific things are off. Tell the model exactly what.

The second paragraph is too formal. Tighten it.

Cut the “we appreciate your patience” line. That’s not how I speak.

Replace the third bullet with something more specific.

The opening sentence is too generic. It could be about any product. Make it specific to ours.

Three rules for critique:

Be specific. “Make it better” is useless. “The third bullet is too long; cut it to one sentence” works.

Be direct. “I think it might be possible to consider revising…” is wasted words. “Rewrite the third bullet” is fine.

Compliment what works. “Keep the structure of paragraph one; rewrite paragraph two.” Otherwise the model may rewrite the whole thing and you lose the parts that were already good.

A useful pattern is to tell the model what to keep, then what to change. “Keep the opening, structure, and tone, but make the middle three sentences sharper and shorter.”

Move 2: Narrow

The first response covered a lot of ground, but the part you actually wanted was one small section. Ask for more of that.

The third option is the one I want to explore. Develop it further: what would it actually look like, what would it cost, and who would need to be involved?

Take the second sentence of paragraph two, “…the regulation may impact existing contracts…”, and expand it into three paragraphs.

Of the five risks you listed, focus on the third. Walk me through how it would actually play out, step by step.

People sometimes repeat the original question with a small tweak when they could instead point to the useful part and ask for a deeper pass. The model already produced the wider draft; you are now narrowing the task.

Move 3: Pivot

The first response is fine, but for the wrong audience or angle.

Now write this for someone who has never worked in finance.

Now rewrite it as if I were sending it to a skeptical CFO instead of a sympathetic colleague.

Take this same set of arguments and structure them as a one-slide summary instead of a memo.

Now do the opposite: the strongest case against everything you just argued.

The pivot move reuses material already in the conversation and reshapes it for a different purpose. Whether that is faster or produces a better result than a clean restart depends on the task, the accumulated context, and the model. Compare both approaches on representative work if the choice matters.

A particularly useful pivot for analytical tasks is: “Now play devil’s advocate. Take your own response and produce the most credible counterargument you can.” The response may expose assumptions or missing evidence, but it is another generated draft, not an independent expert review.

Move 4: Stress-test

The first response is plausible but you are not sure it is right. Make the model defend it or argue against it.

What evidence would change your answer to question 2?

Where in this response are you most uncertain? What might be wrong?

What would the smartest critic of this position say?

If a senior expert read this, where would they push back?

The stress-test move helps reveal weaknesses hidden by polished prose. Ask the model to state assumptions, identify claims that need outside evidence, and construct a credible counterargument. Then assess those outputs yourself rather than treating the model as an independent judge of its own answer.

A useful stress-test for a factual answer is: “Which points depend on facts outside this conversation? For each one, name the primary source I should check and state any assumption you made.” Treat the reply as a starting map for your own verification. A model’s self-reported confidence is not calibrated proof that a claim is true.

The fifth move: verify

The four moves improve the answer. They do not prove the answer is true. Add a final verification pass whenever the output will influence a decision, customer message, policy, code change, or published content.

Use a short checklist:

Iteration can make the draft better; verification decides whether it is safe to use.

The companion workflow card linked from this article gives you the five moves as a one-page habit.

Putting it together: a worked iteration

Let’s run a real example through the four moves. Suppose you start with this prompt:

You are an experienced HR director. Draft a one-page proposal for adopting a four-day work week at my 80-person company. Address the obvious concerns, propose how we’d pilot it, and end with a recommendation.

The first response is a decent draft. Not perfect. What do you do?

Move 1 (Critique):

The “obvious concerns” section is too thin. You covered productivity but skipped customer coverage and how this affects our hourly support team. Add real depth to those two.

Move 2 (Narrow):

The pilot proposal is the part I’ll actually use. Expand it into a four-paragraph plan: who’s in the pilot, how long it lasts, what we measure, and what success criteria look like. Do not invent company facts or thresholds; write [missing] where I need to supply or approve a value.

Move 3 (Pivot):

Now rewrite the same proposal as if the audience were our investor board, not our internal HR team. Different concerns, different language.

Move 4 (Stress-test):

Identify the strongest credible failure case for this proposal. Do not imply personal experience or invent a case study. Separate general reasoning from claims that need an external source.

To make the payoff concrete, here is what Move 2 changes in a representative before-and-after version of the pilot section. Run the prompts yourself for your own variant; the wording and details will differ.

Before: “We recommend piloting the four-day week with one team for a quarter and reviewing the results.”

After: “Pilot with [team or teams to confirm] for [duration to approve]. Before launch, record a baseline for customer coverage, agreed output measures, unplanned overtime, and a voluntary employee pulse measure reviewed for privacy. The pilot owner must approve the success thresholds and stop conditions before the start date; until then, mark them [missing] rather than inventing numbers. Review results at agreed intervals and document any workload shifted to teams outside the pilot.”

Same model, same conversation. The difference is not intelligence. Move 2 told it which part carries the weight, what “expanded” means, and where guessing is not allowed.

After these follow-ups, you should have a proposal whose changes are easier to inspect than a blind restart. Whether it is better depends on the task, the model, the information supplied, and your verification pass.

The patterns that signal “iterate, don’t restart”

Some moments where you should stay in the conversation rather than open a new one:

• The model got most of it right, and you have specific complaints.
• The output is on the right topic but the wrong shape.
• You want to explore alternatives or variations.
• You want to test the answer’s robustness.
• You realised you forgot to mention something important.

Patterns where you should start a new conversation:

• The model has drifted onto something completely different from what you wanted.
• The conversation has accumulated enough material that important instructions are being ignored, contradicted, summarized away, or crowded out.
• You want to test the same prompt with a fresh start to see if you got an outlier response.
• The model keeps offering the same suggestions even when you push back.

A few small habits that help

Stay in one thread while its context still helps. Chat products manage long histories differently: a request may retain earlier turns, use a rolling window, or compact older material. A longer context window also does not guarantee that every detail will be used correctly. Anthropic’s context-window documentation explains the working-memory limit and current compaction behavior; check the documentation for the product you use.

Name the parts. “Paragraph two,” “the third bullet,” “the second option you suggested.” Specific references are easier for the model to act on than vague critique.

Be direct. “That’s too formal. Try again.” “No, that’s worse. Go back to the first version and only change the closing.” Direct editing instructions reduce ambiguity.

Ask the model what it sees. When you have given specific critique and the model keeps missing, try: “What do you think I’m asking for? Restate the goal in your own words.” Its restatement may reveal the mismatch and give you a concrete point to correct.

Keep going after the first answer

Accepting the first response without review can leave important quality work undone. Critique, narrow, pivot, stress-test, and verify form a practical follow-up workflow. The skill is not only writing a better first prompt. It is knowing when another turn will clarify the work, when outside verification is required, and when a clean restart is safer.