Practical Prompting · Intermediate

Lesson 5: Iteration Patterns — Getting from Draft Three to Done

Structured techniques for improving output: single-variable iteration, critique loops, and decomposition.

By the AIVraj Editorial Team 7 min readLast updated July 22, 2026
What you will learn
  • Iterate without losing what already worked
  • Use the model to critique its own draft
  • Break large tasks into chainable steps
In this lesson
  1. Change one variable
  2. The critique loop
  3. Decomposition
  4. Anchoring on a good draft
  5. When to stop

Change one variable

The most common iteration mistake is rewriting the whole prompt after a mediocre result. You then have a different output and no idea which change caused it. Change one thing — the audience, the format, the length limit, the role — and compare. Three controlled iterations beat ten random ones.

The critique loop

Models are noticeably better at evaluating text against explicit criteria than at producing perfect text first time. Use that:

  1. Generate the draft.
  2. In the same thread: "Critique this against these criteria: factual risk, structural clarity, generic sentences, and whether the opening earns the second paragraph. List issues only, no rewrite."
  3. "Now rewrite addressing issues 1, 3 and 4. Leave the rest alone."

Step three matters. Asking for a full rewrite after critique usually regresses the parts that were already good.

Decomposition

When output is consistently mediocre, the task is often too large. Split it. A long article becomes: outline → section drafts → transitions → edit pass. A data task becomes: extract → normalise → analyse → summarise. Each step is short enough to verify, and errors stop compounding silently.

Anchoring on a good draft

Paste your best previous output back in as a style reference: "Match the structure and voice of the example below." This is few-shot prompting applied to your own work, and it is the fastest route to consistency across a series — ten product descriptions, a week of posts, a set of release notes.

When to stop

Stop when the remaining problems are ones you can fix faster by hand. Past that point you are optimising the prompt for its own sake. A realistic split for most professional work is around 70% generated, 30% edited — and that 30% is where the piece stops sounding like everyone else's.

Practise with these tools

Frequently asked questions

Is it better to start a new chat or keep iterating?

Keep iterating while the context is helping. Start fresh when the thread has accumulated rejected drafts, because the model conditions on them and tends to drift back towards earlier attempts.

How many iterations is normal?

Two to four for most professional deliverables. If you are past six, the task probably needs decomposing rather than more prompting.

Next lesson
Lesson 6: Prompting for Images
Continue
More reading: Learning Center · AI Glossary