Gains:
- Understands that the initial output is rarely perfect and that iteration is a normal process
- Can systematically improve an outcome with targeted, univariate feedback
- Knows the logic of turning a successful prompt into a reusable template
The biggest mistake of beginners is to expect the first prompt to give perfect output. When the first output does not meet expectations, they give up saying "AI cannot do this job." However, for experienced users, the first output is not an end, but a beginning. In this unit, we will learn to see prompt writing as a dialogue and improvement process, not a one-time shot. Iteration is the heart of prompt engineering.
What is Iteration and Why is It Normal?
Iteration is taking an output, pointing out what you don't like, correcting the prompt or request, and reproducing it. If the initial output is imperfect, it is not a failure but a natural part of the process. Because most of the time, you only become clear about what you want when you see the first output. The good user treats the initial output as a "draft" and whittles it towards the goal.
There are two types of iterations:
- In-conversation correction: Proceeding in the same conversation with feedback such as "make this shorter", "remove the third item". It is fast; the model remembers the previous output.
- Rewriting the prompt: Applying what you have learned to the initial prompt and starting from scratch with a better instruction. It is valuable for works that will be reused.
Giving Targeted Feedback
The secret of iteration is that the feedback is targeted. Saying "I don't like it, write again" does not direct the model; You will probably get a similar output. Instead, say exactly what you would change and how:
- Weak: "This is not good."
- Strong: "The introduction is too long; reduce the first paragraph to one sentence. The tone is too formal; make it a little more friendly. Remove the claim from the second bullet point because we cannot verify it."
Univariate optimization
Make a scientific habit: change one thing at a time. If you change tone, length, and structure all at the same time, you won't be able to tell which change took effect, whether the output improves or deteriorates. Going one-variable allows you to learn what works and make it stick.
Iteration Loop
step
what do you do
1. Produce
Get the output with the first prompt
2. Evaluate
What you liked, what you didn't like - write concretely
3. One change
Make a single fix that targets one issue
4. Regenerate
See the effect of the change
5. Compare
Has he recovered? If yes keep it, if no take it back
6. Fixing
Save the prompt that gives good results to the template
Four Copiable Templates
1) Targeted correction:
The previous output was fine but change:- [single concrete change 1]- [single concrete change 2]Keep everything else the same.
2) Critique the output:
Critique the text you just produced with an editor's eye. Find 3 weak points and suggest a concrete correction for each. Then write the corrected version.
3) Variant generation (A/B):
Make 2 different versions of this text: A) More formal and shorter. B) More friendly and with examples. Give both; Tell me in 1 sentence which one is more suitable in which situation.
4) Make Prompt heal himself:
I use the following prompt, but the result is enough [problem]. Suggest how I can write this prompt more clearly and effectively; write the improved prompt and explain what you changed and why. Prompt: [current prompt]
Weak Prompt / Strong Prompt (as feedback)
Poor feedback:
It didn't work, try again.
Strong feedback:
The content is correct but there are three problems:1) 210 words; Reduce to 120.2) Second paragraph is technical; simplify the jargon.3) The closing is weak; End with a clear next step sentence. Keep everything else.
The second feedback tells the model exactly where to touch; The result gets closer to the target in one turn.
Three Mini Cases
Case 1 — Content production. A marketer tried a promotional copy by saying "I don't like it, write it again" 6 times and never got any further. When he changed his approach and asked for one concrete change in each round (first the length, then the tone, then the closing), he achieved a publishable text in 3 rounds. Total time reduced to less than half.
Case 2 — Report. One analyst saw that the first summary was too technical and said, "the manager won't understand this, translate the technical terms into everyday language, keep the length." Thanks to univariate correction, the report was both simplified and did not lose its scope; There was no need for a second complete rewrite.
Case 3 — Templating. A support team was correcting the same type of response iteratively, time after time. Finally, they recorded the ideal prompt in a template; They now use the template with minor changes instead of iterating from scratch with each new case. This illustrates the point of iteration: make it good once, then use it again.
Tip: Ask the model to critique her own output (“find 3 weaknesses of this text”). The model often realizes its own error and corrects it; this does a pre-editing on your behalf.
Caution: Iteration has a limit. If after 5-6 rounds you are still not getting close to the goal, the problem is not in small corrections, but in the basic approach: you may have set up the role, context or task incorrectly from the beginning. At that point, stop patching and redesign the prompt from scratch.
Two Paths of Iteration: Conversation or Rewriting?
Knowing which way is more efficient while iterating saves time. The rule is: for a one-time job, it's quickest to fix it in chat; because the model remembers the previous output and you just tell the difference. But for a job that you will repeat, it is more valuable to write what you have learned into the initial prompt and produce a clean template; because next time you will get a good output without any need for iteration.
In practice, it is most efficient to combine the two: first discover the correct instruction by trying it quickly in the chat, then pour the instruction you find into a clean prompt and save it in the library. In this way, you can both progress quickly and turn your labor into a permanent asset.
Viewing iteration as a learning record
Each round of iteration teaches you something: which addition improved the output, which constraint was unnecessary, what the model got wrong. Writing down these learnings creates a personal "what works" guide over time. For example, inferences such as "the 'add a comment' constraint is a must in this type of summaries", "3 examples are sufficient for title production, 5 are unnecessary" will enable you to write the correct prompt on the first try in the next similar job. This knowledge is what really differentiates the experienced user from the beginner: they start each business not from scratch, but armed with the lessons of previous iterations.
Making the model an iteration partner
You don't have to run the iteration alone. You can ask the model, "How can I bring this output closer to my goal, give me 3 concrete suggestions." The model often accurately diagnoses weaknesses in its output and suggests ways to fix them. This transforms iteration from “trial and error” into targeted collaboration; You tell us what you want, the model tells you how to get there.
Common mistakes
- Sending the same prompt again. Trying again without changing; the result does not change.
- Vague feedback. Saying "I don't like it" and not saying what you will fix.
- Changing many things at once. Not learning what works.
- Not recording the good prompt. Doing the same iteration over again each time.
- Infinite iteration. Trying to fix a fundamentally wrong prompt by patching it.
In summary
- Initial output is rarely perfect; Iteration is not a failure, it is a natural part of the process.
- Effective iteration proceeds with targeted, not vague, feedback.
- Change one thing per round; so you learn what works.
- You can pre-edit by asking the model to critique her own output.
- Save the prompt that works well in the template; The goal of iteration is to produce a reusable asset.
Application task
Take a printout and consciously go through three rounds of iteration: changing only one thing per round (first the length, then the tone, then the closing), and record each round. After the third round, determine which change made the most difference and save the final prompt as a reusable template.
checklist
- [ ] I view the initial output as a draft, not as finished work.
- [ ] I give my feedback concrete and targeted.
- [ ] I make one change per round.
- [ ] I can ask the model to critique its own output.
- [ ] I save the prompts that give good results to the template.