What it is
Most good results take a few rounds. You ask for the weekly update, then say “shorter”, “put a date on anything that slipped”, “no ticket numbers”, “drop the excited tone”. By the fifth message the answer is right. Next week you start again from the same short request and make the same five corrections.
Turning the chat into a prompt ends that loop. While the conversation is still open, ask the model to write a single prompt that would produce the final version from the start. Every correction becomes a rule. Every fact that will change next time, such as this week’s notes, becomes a labeled blank. Facts that stay the same, such as your refund policy, stay written out.
The chat already holds what the model needs: your first request, each correction and the version you accepted. Writing the prompt is mostly summarizing, which models do well, and you get the summary while it is fresh.
- 1Get one good resultCorrect the model in the chat until the answer is right.
- 2Ask for the promptCorrections become rules; facts that change become blanks.
- 3Test it on a new caseA fresh chat, different input, no extra help.
- 4Save itKeep it, and each later version, in My Prompts.
When it helps
Use it after any chat that took several corrections for a task you will do again: weekly updates, customer replies, report sections, meeting notes. Use it before you hand a task to a colleague, so they get the prompt and not a description of what you usually say. And use it to set up a standing assistant, such as a Custom GPT, a Claude Project or a Gem, where the prompt becomes the instructions.
It is also a way to learn what worked. Read the rules the model lists: they are the parts of the brief you left out the first time, and the next prompt you write from scratch can include them.
When it does not
A one-off task does not need a reusable prompt. Keep the answer and move on.
One-case details can turn into rules. If you told the model to mention the delayed March order, the prompt may now tell it to mention delays in every update. Read the draft for anything that belonged only to this case.
Long chats lose things. In a conversation with dozens of turns, an early correction can be missed or outweighed by a later one. Ask the model to list every correction first, check the list, then ask for the prompt.
The model knows only what is in the chat. If you fixed something by editing the answer yourself, it never saw that fix. Add it.
And the prompt it writes is untested. It reproduced one good answer in a chat that had all the context. Run it on a fresh case in a new chat before you rely on it, and expect to trim it: generated prompts tend to run long.
Before and after
In each pair, Before is the first message of a chat that took several rounds, and After is the prompt the model wrote at the end. The scores are the Prompt Grader’s free estimate with Written for set to Any chat model. A score counts what the prompt states; it does not grade the answer.
Write my weekly update for the team from these notes. """ [notes] """
Notes for this week: """ [paste this week's notes] """ Write my weekly update for the product team from the notes above. The team reads it on Monday morning to see what shipped, what slipped and what needs their input. Format: three headings, "Shipped", "Slipped" and "Need your input", each with up to four bullets of one sentence. Put a date on anything that slipped. Leave out internal ticket numbers and anything marked "private". Write in plain sentences, no exclamation marks and no "excited to share". If the notes do not say why something slipped, write [REASON?] instead of inventing one.
The first message got a long, upbeat summary that needed five corrections. The prompt carries all five, so next Monday’s update comes back in the agreed shape on the first try, with a visible [REASON?] where your notes are thin.
Help me answer this refund request. """ [customer email] """
Customer email: """ [paste the customer's email] """ Draft a reply to the refund request above for our support team to send. Our policy: full refund within 30 days of purchase; after 30 days, store credit only; no refunds on gift cards. Decide which rule applies from the purchase date in the email, then write a reply of under 120 words that states the decision in the first sentence, explains it in one sentence, and ends with what happens next. Tone: friendly and direct. Do not offer exceptions to the policy. If the email gives no purchase date, ask the customer for it instead of deciding.
In the chat, you had to explain the policy, then stop the model offering an exception, then ask for the decision first. The prompt has the policy written in and those rules attached, so a colleague can paste in any refund email and get a reply that follows policy.
The line to add
Send this at the end of the chat, once the answer is right:
That last version is what I wanted. Write a reusable prompt that would get a result like it in one go next time.
- Turn every correction I made in this chat into a rule.
- Replace facts that change each time with labeled blanks, such as {{notes}}.
- Keep facts that stay the same, such as our policy, written out.
- Leave out anything that applied only to this one case.
- Keep it under {{number}} words, and return only the prompt.
For a long chat, ask first: “List every correction I made in this chat, one line each.” Check the list, then ask for the prompt. Paste the result into a new chat with a different case to test it, and save the version that works to My Prompts, which keeps every version as you refine it.
In ChatGPT, Claude, Gemini and Copilot
The request is plain text, so it works the same way in all four, as long as you send it in the same conversation. A new chat does not have your corrections.
Where you keep the result differs. Microsoft 365 Copilot can save a prompt you have run: hover over it, choose Save prompt, and it goes to Your saved prompts in Copilot Prompt Lab, visible only to you (Save your favorite prompts). That saves the prompt as you typed it, so run the rewritten prompt once, then save that one.
In ChatGPT, Claude and Gemini, a prompt you use daily can become the instructions of a Custom GPT, a Claude Project or a Gem, so you only paste the part that changes. The custom instructions guide covers what belongs there and what belongs in each message.
Key takeaways
- Ask for the prompt in the same chat, while the corrections are there.
- Corrections become rules; facts that change become labeled blanks.
- Read the draft for details that belonged only to this one case.
- In a long chat, have it list the corrections first and check them.
- Test it on a fresh case in a new chat, then save it.
Questions
How do I save a prompt that worked in ChatGPT?
Ask ChatGPT, in the same chat, to write the reusable prompt that would get the final answer in one go, with blanks for what changes. Test it in a new chat, then keep it somewhere you will find it: My Prompts, a shared document, or the instructions of a Custom GPT if you use it daily.
Can ChatGPT or Claude write a prompt from our conversation?
Yes. Both can read the whole conversation and summarize your corrections into one prompt. Ask for every correction as a rule and every changing fact as a labeled blank, then check the result for details that applied only to this case. In a very long chat, ask for the list of corrections first.
How do I make a reusable prompt template?
Start from a prompt that already worked. Replace everything that changes from one use to the next, such as the notes, the customer email or the date, with a labeled blank like {{notes}}, and keep fixed facts such as policies written out. Test it on two or three different cases before you share it.
Prompts to use with this
- Write worked examples that fade into practice problemsYou are teaching a procedure, such as solving equations or calculating a dose, and need step-by-step examples that lead into independent practice.