Iterate: when the first answer is wrong

When an AI’s first answer is wrong, work out what exactly is wrong before you type anything back. Keep the parts that worked, ask for one small change, and fix the instructions you gave instead of arguing about tone for seven rounds while a stale number sits in the draft.

Say your coworker in accounts payable sends you a vendor email drafted with ChatGPT. You have a spreadsheet open in the next tab, and it says the Q2 overage for Northline Logistics is $18,640. The draft says $14,200, a figure from a 2025 rate PDF that was attached to the first prompt “for context.” You have typed “fix the tone” seven times. The greeting got warmer each time, but the dollar amount never moved, and the seventh version sounds so human that you almost send it.

First answers are useful and specifically wrong

A first reply can look finished, with complete sentences, a subject line, and a polite closing. That is the trap. Models are good at the shape of an email, a study outline, or a shopping comparison, and they are sloppy with the one fact you will get in trouble for. That fact is usually a dollar amount, a date, or the name of someone who should not be copied.

A draft is useful when the job is already in it. In the Northline email, “please issue a credit memo” and the Friday deadline were keepers. A draft is specifically wrong when it holds a detail that would be false if you sent it. The $14,200 was false, and the stock greeting was the wrong voice for a vendor’s payables team.

OpenAI’s prompting page says the first prompt does not need to be perfect: you review the result, then ask for the specific change you want. Anthropic and Google give the same advice, which is to iterate. None of them tell you to type “make it more professional” seven times, because that names a mood and not an error. Check numbers first, names second, and tone last, since a warm email with the wrong money is still a bad email. Here the true figure lived in cell E14 of the file Q2_vendor_recon_v3.xlsx, and the chat was never going to know that.

Name the error before you type

Before you send a follow-up, describe the failure in one sentence you could read aloud. For example, “The opener is a stock greeting I never use,” or “The overage is $18,640, not $14,200.” If you cannot quote the bad line or the bad number, you are not iterating, you are rolling the dice.

Four kinds of wrong cover almost every first draft that feels off, and each one calls for a different next prompt. Try not to mix them in one turn, because that is how a tone fix ends up rewriting a number you had already corrected.

All seven of your turns in the Northline story were about tone. The fact error stayed in the draft because you never pointed at it, and the model had no reason to drop $14,200 when you had given that number a home by attaching the old PDF.

Rule of thumb: If you cannot quote the bad line or the bad number, you are not iterating yet, because “fix the tone” is a shrug.

Keep what worked, then ask smaller

A follow-up that says “rewrite this” throws away the good half of the draft. The model does not know which paragraph you liked unless you say so, so quote the keeper. Try “Keep the credit-memo ask and the Friday deadline, and do not touch those two sentences.” That puts a fence around work that already passed.

Add one new rule per turn. If you ban the greeting, lock $18,640, cap the length at 90 words, and forbid legal language all at once, the next draft will obey two of them and invent a fifth problem. You will not know which instruction landed. Your manager’s real ask is one job, which is to get the right invoice and the right money in front of Northline before they close their books.

Iterating builds on the first-prompt kit from the earlier post on context, constraints, and examples, applied to a draft you can now see. The context is the draft plus the sheet, and the constraint is the one new fence. The example is two lines you would really send: “Hi Northline AP, Invoice 8841 shows an overage of $18,640 against our Q2 recon. Please issue a credit memo by Friday.” Paste that in, rather than describing what “professional” means.

Regenerate is a coin flip. A named follow-up is a steer, so ask for a local edit. Tell it to change only the opener, or to replace $14,200 with $18,640 everywhere. If the same error comes back, shrink the ask again and say “return the email with this opener deleted and nothing else changed.” A thread that keeps restoring $14,200 may still have the 2025 PDF in view, which means the cleanest fix is a new chat with the recon numbers only.

A five-step iterate script you can paste

Copy this into the chat after a draft lands and fill in the brackets. Send it as one message, and then read the reply the way you read the first draft, with the numbers first.

Iterate script (paste, then fill the brackets)

1. NAME THE ERROR
The last draft is wrong in this specific way:
"[quote the bad sentence or the bad number]"
Do not rewrite the whole piece.

2. KEEP WHAT WORKED
Keep these parts unchanged:
- [subject / paragraph 2 / the Friday ask / the table]
Do not drop them to improve the rest.

3. ADD ONE CONSTRAINT
Add this single new rule: [ban the stock greeting / use $18,640 only / 90 words / no legal language]
Do not add a second rule in this turn.

4. SMALLER ASK
Change only: [the opener / the dollar figure / the sign-off]
Return the full draft with that local edit, not a new essay.

5. STOP CHECK
Count my named fixes on this error: [1 / 2 / 3]
If this is the third time the same error remains, say so in one line and stop.
I will edit it myself or start a new chat with a tighter brief.

The script makes you quote, protect, and count. In the story you skipped the stop check, so turn seven felt like progress. The count of fact fixes was zero, and the count of greeting fixes was seven, none of them named.

A worked example: seven tone prompts, one stale number

Here are the facts of the case. Invoice 8841 has a true overage of $18,640 from cell E14, and the stale $14,200 came from the 2025 PDF. The audience is Northline’s payables team, not the salesperson, and the goal is a credit memo by Friday in 90 words with no legal language. Seven tone turns went like this.

The turn that would have ended it was not another mood word. It would have said: keep the credit-memo ask and the Friday deadline, delete the opener, and start with invoice 8841. Use $18,640 from the recon sheet, and do not mention legal action. Keep it to 90 words, addressed to payables only. If you want one rule per turn, lock $18,640 first and ban the greeting second, because two short turns beat seven long ones.

Arguing with the model versus fixing the brief

Arguing looks like a stack of adjectives with no new source: warmer, less stiff, more human, not that human, more like how I write. The model keeps guessing because you have not shown a sample or banned a line. These taste fights burn twenty minutes and still ship the stock greeting in a new costume.

Fixing the brief looks like new information entering the chat. You paste E14, you paste two lines you sent last month, and you name the audience as payables and not the account executive. You also say which file is stale, for example “Ignore the 2025 PDF, it is last year’s rates.” A follow-up is allowed to correct a thin first prompt or a wrong file. Here is a quick test: did this turn add a fact, a quote, a ban, or a format limit? If it only added a feeling word, you are arguing, so paste a sample or edit the opener yourself. If the model is stuck on a number it saw earlier, start a new chat with the recon figures and without the 2025 PDF.

When to stop iterating

Stop when the same named error is still there on the third try, because a fourth “fix the tone” will not discover cell E14. Stop when you are in a tone war and no new fact has entered the chat. Stop when a dollar amount, date, or name cannot be pointed at in a sheet you trust, and do not send a figure you did not check. Stop when the remaining work is a delete or a paste you already have on screen.

Twenty minutes is a ceiling for one short email. Set a timer. The seven turns in the story ran longer, and the $14,200 never got a glance. Polishing tone around an unchecked number is decoration. Check the money. The post on when to stop prompting and do the task yourself maps out when to close the chat. If step 5 of the script trips, you are done prompting on that error.

Common mistakes

• Typing “fix the tone” or “make it more professional” without quoting the bad line.
• Leaving a stale PDF in the thread, then wondering why the old number keeps coming back.
• Rewriting the whole draft and losing the paragraph that already worked.
• Stacking five new rules in one follow-up, then being unable to tell which one landed.
• Using Regenerate instead of a named local edit.
• Checking tone for seven turns and never checking the money.

How to practice this week

Take one real draft, such as a vendor email, a study outline, or a short plan, and run the five-step script once. Quote one error, keep one keeper, and add one constraint. If the same error survives three named tries, stop and edit it yourself. The next post covers study, work email, plans, and shopping research, and the first-prompt kit is in context, constraints, and examples. For more writing practice, see AI for writing and questions and Learn.

Quick recap

• First drafts are often useful in shape and wrong in the fact you will get blamed for.
• Name the error (facts, tone, format, or a missing constraint) and quote the bad line.
• Keep what worked, add one constraint, and ask for a smaller edit instead of a new essay.
• If the same error remains on try three, stop, then paste a better brief or edit it yourself.
• The $14,200 survived seven tone prompts because nobody checked the sheet.

Series notes

This is Part 3 of Prompting for everyone (iterate). Previous: Study, email, plans, shopping. Next: When to stop prompting.

Sources

• OpenAI: Prompting (ChatGPT Learn) (first prompt need not be perfect; follow-ups name the specific change)
• OpenAI: Prompt engineering (iterate on prompts; review after each change)
• Anthropic: Prompt engineering overview (clear instructions, then refine)
• Google: Gemini prompting strategies (design, test, iterate)
• AMS: Prompting for everyone
• AMS: Give context, constraints, and examples (the first-prompt kit this part iterates on)

Merch on Etsy

Open store →

You're On Mute Coffee Mug View on Etsy → Apache Hadoop T-Shirt View on Etsy → Standard Deviation White Ceramic Mug View on Etsy → Mean Squared Error Shirt View on Etsy →