Prompting 105: iterate เหมือน reviewerPrompting 105: iterate like a reviewer

The first answer is a draft.

This is one of the most important AI literacy habits. If you treat the first output as final, you let the model decide the quality bar.

Better:

Use the first answer to see what the model understood, then steer it.

Use review prompts

After AI gives an answer, ask it to review against criteria.

Review your answer against these criteria:
- clear for a non-technical reader
- no unsupported claims
- specific next steps
- under 150 words

List issues first. Then provide a revised version.

This separates judgment from rewriting.

Ask for alternatives

One output can hide better options.

Give me 3 versions:
1. direct and concise
2. warmer and more reassuring
3. more formal for senior management

After the 3 versions, recommend which one fits best and why.

Alternatives help you choose rather than accept.

Use a rubric

A rubric turns taste into criteria.

Score this draft from 1-5 on:
- clarity
- accuracy
- usefulness
- tone
- actionability

For any score below 4, suggest a fix.

This is useful for emails, slide outlines, summaries, and plans.

Ask what is missing

If the output feels generic, ask:

What important information is missing from my prompt?
Ask up to 5 questions that would help you improve the answer.

This teaches you how to prompt better next time.

Compare versions

When you have two drafts, do not ask "which is better?" Ask for specific comparison.

Compare Draft A and Draft B for this audience: [audience].

Evaluate:
- clarity
- trust
- risk of misunderstanding
- likelihood that the reader acts

Then recommend one version and explain the tradeoff.

Good prompting often means becoming a better editor.

Know when to restart

Sometimes follow-up prompts cannot fix a bad start. Restart when:

• the original task was unclear
• the context was wrong
• the model invented facts
• the output format is completely wrong
• you changed the goal halfway through

When that happens, write a cleaner prompt from scratch.

Practice: second prompt

Ask AI to draft something small: an email, summary, or meeting agenda.

Do not edit it yourself first.

Send this second prompt:

Review this as if you are my careful editor.

Goal:
[what this output should achieve]

Audience:
[who will read it]

Criteria:
- clear
- specific
- no unsupported claims
- easy to act on

First list problems. Then rewrite.

Compare the first answer with the revised answer.

References

• Anthropic Prompt Engineering Overview recommends defining success criteria and empirical tests before prompt tuning.
• OpenAI Prompt Guidance emphasizes outcome-first prompts, success criteria, constraints, evidence, and stopping conditions for newer models.
• OpenAI GPT-5.2 Prompting Guide emphasizes prompt iteration with evals, scope control, and concrete output constraints.
• Yang et al., 2023: Large Language Models as Optimizers explores using LLMs to optimize prompts, reinforcing that prompt quality can be iterated rather than guessed once.