Refinement
Iterate
Prompting is a conversation. The first answer is often a draft, not the finish line.
Start with an initial prompt, review the response
OpenAI Help Center
Key Facts
- A follow-up prompt should name the miss, add the missing detail, and ask for a targeted revision.
- Do not restart from scratch when only tone, length, format, or one assumption needs repair.
- Important claims still need verification even after a better follow-up answer.
Beginners often blame themselves when the first answer is not right. More often, the prompt simply did not include enough information yet. Official guidance from OpenAI, Google, and Microsoft all treats follow-up as normal: ask, review, refine, and keep the conversation going.
The best follow-up prompts are specific. Do not say only "try again." Say what missed, what matters now, and what the revised answer should do differently.
The repair loop
- Name the miss: "This is too generic," "The tone is too formal," or "You assumed we have a budget."
- Add the missing detail: provide the audience, policy, example, deadline, source text, or constraint that was absent.
- Ask for a targeted revision: request only the change you need instead of restarting the whole task.
- Verify: check important facts, citations, calculations, names, dates, and claims before using the answer.
Too generic
Weak
This is too generic. Make it better.Better
This is too generic for our audience.
Revise it for first-time managers at a 40-person startup.
Keep the same structure, but add concrete examples involving one-on-ones, unclear ownership, and missed deadlines.
Avoid corporate buzzwords.The better follow-up names the audience, preserves what should stay, adds concrete situations, and blocks an unwanted style.
Wrong format
Weak
I wanted a table.Better
Convert the answer into a table with these columns:
Task | Owner | Why it matters | Deadline | Risk if skipped
Keep each cell under 18 words. Do not add new tasks.The better follow-up gives the exact table shape and prevents new invented content.
Unsupported claims
Weak
Are you sure?Better
Review your answer for unsupported claims.
For each claim, mark one of:
- Supported by the source I gave
- Reasonable inference
- Needs verification
Do not add new claims.The better follow-up creates a verification pass instead of asking for vague confidence.
When the answer is wrong
If the model makes a factual error, correct it directly and provide the source if you have one. Then ask it to revise the answer using the correction. For important facts, do not stop there. Check the source yourself. OpenAI's accuracy guidance warns that models can sound confident even when wrong, including fabricated quotes, studies, citations, or references.
When the answer is too long
Give a hard shape. "Shorter" is subjective. "Five bullets, each under 12 words" is clear. "One paragraph under 90 words" is clear. "Only the revised email, no explanation" is clear. These constraints are especially useful for emails, social posts, executive summaries, and slide copy.
When the tone is wrong
Tone words help, but examples help more. If the answer sounds too stiff, paste one sentence that has the tone you want. If the answer is too casual, say who will read it and what relationship you have with them. A tone prompt should describe the social situation, not just the mood.
Save the improved prompt
The prompt you end up with after two follow-ups is often better than the one you started with. Save it as a recipe. Next time, begin with the improved version and change the context. That is how everyday prompting gets faster without becoming mechanical.