What is iterative prompting?

Few prompts are right the first time, even for experienced users. Iterative prompting accepts that and turns it into a method. You send a reasonable first request, treat the reply as information about what the model understood, and use your next message to close the gap. Because ChatGPT keeps the whole conversation in view, each round builds on the last instead of starting over.

A Beginner's Guide to ChatGPT describes this as a cycle: start general, look at the answer, decide where it falls short, then adjust by adding context, clarifying what you want or giving more guidance, and repeat. The useful part to practise is the middle step. A weak answer is a symptom, and each symptom points to a specific missing ingredient.

Reading the symptoms

OpenAI's prompting guide gives the same advice from the developer side: be specific about the outcome, length, format and style, and say what to do rather than only what to avoid.

A worked example in three rounds

A volunteer at a seniors' club has been asked to give a 10-minute talk on avoiding online scams. Here is how one conversation could go. The replies are described in summary and are illustrative.

Round 1.

Help me prepare a talk about online scams.

The reply covers phishing, ransomware, identity theft and data breaches, with security advice such as two-factor authentication and password managers. It is accurate, long, and pitched at office workers.

Diagnosis: the model did not know the audience, the length of the talk, or the goal. It answered the average version of the question.

Round 2.

The audience is about 25 members of a seniors' club, mostly in their 70s. Many use email and Facebook but are not confident online. The talk is 10 minutes. The goal is that they can recognize three common scams and know what to do. Please replace your outline with one for this group.

The reply now picks three scams relevant to the group, such as a fake message from a bank, a grandchild-in-trouble call and a fake prize, with warning signs for each. But it runs to two pages, and some advice uses words like "browser extension" and "spoofing."

Diagnosis: audience and purpose are fixed. What is missing now is format and level of language.

Round 3.

Good choice of scams. Now:
- Fit it to 10 minutes: an opening of 1 minute, about 2.5 minutes per scam, a 1-minute close.
- For each scam: one short realistic example, three warning signs, and one thing to do.
- No technical terms. If you must use one, explain it in a few words.
- End with a one-page handout in large-print style, five lines maximum.

The reply now fits the talk. One step remains, and it belongs to you: check the facts. Any phone number, organization or statistic the model includes must be confirmed before you print the handout, because ChatGPT can write wrong details in a confident tone. Official government and police websites are the right places to confirm what to tell people.

Why iteration works

Two properties of ChatGPT make this approach effective. First, it follows the conversation, so a short correction like "shorter, and no jargon" is applied to the existing draft. Second, as OpenAI noted when it launched ChatGPT, the model is sensitive to how a request is phrased, and it usually guesses when a request is ambiguous rather than asking. Each round removes one ambiguity and so removes one guess.

How do you fix a weak first answer?

Name the problem before you write the fix. "That's not what I wanted" gives the model nothing to work with; it will produce another guess. Compare the reply with what you needed, find the single biggest gap using the table above, and say exactly that: "This is for beginners, not professionals," or "I need five bullet points, not an essay."

Keep what worked. Starting your follow-up with "The three scams are right; keep them" protects the good parts while you fix the rest. Without it, the model may rewrite everything and lose them.

Change one or two things per round. If you fix the audience, the length and the tone at once and the answer gets worse, you will not know which change did it. Small steps make the cause visible.

Know when to start fresh. After many rounds, a conversation can carry old instructions that pull against new ones. When that happens, ask ChatGPT to summarize the final requirements as a single prompt, open a new chat and paste it in. You end up with a clean request, and a prompt worth saving for the next time you need the same kind of answer.