BeginnerLesson 8 of 305 min read

Iterating: improving a prompt step by step

Learn the improvement loop, how to use follow-ups, keep a version log, ask the AI to improve your prompt and know when to stop.

In this lesson you will

  • Run the improve loop
  • Choose between follow-ups and prompt edits
  • Keep a version log
  • Know when to stop

About 16 minutes

Your first prompt is a draft

Nadia asks for a product description and receives something decent but too long, with a claim she cannot support. She almost starts over. Instead she replies: Shorter, remove the claim about allergies, and start with the scent. The next version is nearly perfect. Professionals rarely write the final prompt in one go. They write, look, adjust and try again. This lesson shows you how to iterate quickly and on purpose.

The improvement loop

  1. Run your prompt and read the answer carefully.
  2. Name the gap: what exactly is wrong or missing?
  3. Choose the smallest fix: a sentence of context, a format line, an example.
  4. Run again and compare with the last version.
  5. Keep what works and save the winning version.

Changing one thing at a time tells you what helped. Changing five things at once may improve the answer, but you will not know why, and you cannot repeat it.

Two ways to improve: edit the prompt or follow up

Method How Best when
Follow-up message Reply in the same chat: ‘Shorter, and start with the scent’ Quick adjustments to one answer
Edit and rerun the original prompt Improve the first message and run again You want a clean, reusable prompt
Start a fresh chat Paste your improved prompt into a new conversation The chat became confused or too long

A habit that pays off: when a follow-up fixes the result, copy the fix into your original prompt so next time you get it first time.

Useful follow-up phrases

  • “Shorter: keep only the three most important points.”
  • “More specific: add one concrete example for each point.”
  • “Same content, simpler words for a 12-year-old.”
  • “Give me three alternative versions with different tones.”
  • “What did you assume that I did not tell you?”
  • “Check your answer against my original requirements and list anything you missed.”

Keep a version log

Treat prompts like drafts of a document. A tiny log makes improvement visible. This program scores three versions of a prompt on five simple checks, as an illustration of tracking progress. The scores are crude, but the habit of comparing versions is genuine. Output is real:

versions = {
    "v1": "Write about our shop.",
    "v2": "Write a 50-word welcome message for our candle shop's website.",
    "v3": "Write a 50-word welcome message for our candle shop's website. Audience: first-time visitors. Tone: warm, no exclamation marks. Mention hand-poured soy wax and free delivery over Rs 3,000. End with a call to browse the collection.",
}
checks = {
    "has number": lambda t: any(c.isdigit() for c in t),
    "names audience": lambda t: "audience" in t.lower(),
    "sets tone": lambda t: "tone" in t.lower(),
    "gives facts": lambda t: "soy" in t.lower() or "delivery" in t.lower(),
    "asks for an ending": lambda t: "end with" in t.lower(),
}
for name, text in versions.items():
    score = sum(1 for f in checks.values() if f(text))
    print(name, score, "/", len(checks), "|", len(text.split()), "words")
v1 0 / 5 | 4 words
v2 1 / 5 | 10 words
v3 5 / 5 | 36 words

Version 3 scores best because it names the audience, tone, facts and ending. In real work you would also read the actual answers, because a prompt can score well on a checklist and still produce a weak result.

Ask the AI to improve your prompt

Here is a prompt I wrote and the answer I got. The answer was too generic. Suggest an improved version of my prompt that would give a more specific answer, and explain in one line each what you changed.

My prompt: [paste]
The answer: [paste]

Treat the suggestion as a draft. Run it, compare and decide. The AI is good at spotting missing context but cannot see what is in your head.

Knowing when to stop

  • The answer is usable with only light editing.
  • Further changes do not improve it, only change it.
  • You are spending more time prompting than writing it yourself would take. A small task may be quicker by hand.
  • The model keeps making the same mistake. Try a fresh chat, add an example, or split the task.

Troubleshooting repeated failures

Pattern Try
Keeps ignoring one rule Move that rule to the end, make it a short separate line, and give an example
Gets worse over a long chat Start a new chat with the best prompt and key facts
Invents details Say ‘use only the facts below; if missing, write UNKNOWN’
Style drifts Paste a sample and say ‘match this’
Common mistake

Scrapping a prompt after one bad answer. Read the answer, find the specific gap and fix that. One edit often turns a poor prompt into a good one.

Save the winners

Whenever a prompt works well, paste it into a notes file with a title and a date. After a month you will have a library of tested prompts, which is exactly what the project lesson builds.

Check what you learned

Five quick questions. Pick an answer and you will see straight away whether it is right and why.

Why change one thing at a time?

  1. It is faster to type
  2. The AI prefers it
  3. It avoids errors forever
  4. You learn what helped

Single changes show cause and effect.

What is the first step of the loop?

  1. Delete the chat
  2. Run the prompt and read the answer
  3. Change everything
  4. Buy a plan

Read before you change.

What is a good habit after a follow-up fixes the answer?

  1. Forget it
  2. Delete the answer
  3. Add random words
  4. Copy the fix into the original prompt

Next time you get it first time.

What can you ask to reveal gaps?

  1. What did you assume that I did not tell you?
  2. Who are you?
  3. What time is it?
  4. Can you sing?

It surfaces hidden assumptions.

When should you stop iterating?

  1. Never
  2. After one try
  3. When the answer is usable with light editing
  4. When it gets longer

Stop when more changes do not improve it.

Summary

  • Your first prompt is a draft.
  • Name the gap, make the smallest fix, run again.
  • Copy successful fixes into the original prompt.
  • Save winning prompts and stop when answers are usable.

Next lesson

Next: common mistakes and good habits, including accuracy and safety.

Questions

How many iterations is normal?

Two to four for most tasks. Complex jobs may take more.

Should I iterate in the same chat?

For small adjustments yes. For a clean start paste your improved prompt into a new chat.

Can the AI fix its own mistakes?

Often yes when you point to the specific problem, but you must still check the corrected version.

Progress is saved in this browser.