Start
Menu

Saving good prompts · for your AI

Saving good prompts

---
id: keeping-an-ask
title: Saving good prompts
topic: usage
level: 1
builds-on: [a-good-ask]
modes: [day-pack, rucksack]
moment: "an ask has just worked after a correction or two, they retype an ask they have used before, or they say they did this last week and cannot remember how they asked"
keywords: [saved prompt, prompt that worked, reuse an ask, notes app, prompt library, single ask, saved ask, custom GPT, Gems, paste again, single prompt, a note you own, reuse a prompt, private details, labels, line beside the prompt, retest, worse replies]
related: [a-good-ask, custom-instructions, your-boundary]
checked: 2026-10-27
---

What it is

A prompt is the message you send your AI app. A prompt that works is often built up over several messages and corrections. Your AI app can rewrite the wording that worked as a single prompt. A single prompt would have got the reply you wanted without the earlier chat.

Save the single prompt in a note you own. That means a note in your notes app or a document, kept by you, not an AI app. Beside the saved prompt, write one line on what the prompt is for and what a good reply looks like.

Suppose a reply to a customer took three tries, and you saved the single prompt, with "the customer" for the name. Two weeks later you type a fresh prompt from memory and the reply is stiff. So you paste the saved prompt into any AI app and add the name.

Tips

  • When a prompt finally works, ask your AI app to rewrite it as a single prompt, because the wording that worked is often a first try plus your corrections.
  • Save each single prompt in a note you own with the line beside it, because the note works in any app and that line is what you check replies against.
  • When a saved prompt gives worse replies, paste it into a new chat, compare the reply with the line beside it, and reword it in your note, because apps change.

Try it

  1. In a chat from this week where a job took a few tries and you liked the final reply, send this in that chat: "Write one prompt that would have got this reply first time, with no need for the earlier messages."
  2. Copy that prompt into a note you own, then swap any names or numbers in it for labels, plain words such as "the customer", so private details stay out of the note.
  3. Under it, write the line: what the prompt is for, and what a good reply looks like.
  4. Open a new chat, paste the prompt from the note with the real names put back, and read the reply against the one you liked.

The new reply should be close to the one you liked, not identical, because your AI app words its reply a little differently each time you send the same prompt. If the two replies look far apart, send the same prompt once more before you judge it.

If the new reply leaves out something you liked, the earlier messages carried it, not the prompt. Add the missing part to the prompt in your note, then paste again.

A reply close to the one you liked means the saved prompt works from the note alone, without the old chat.

No chat from this week? Use a job you repeat, such as a monthly customer reply, correct it once, then follow the steps.

Risks

  • A saved prompt can carry private details, such as a customer's name, and the company that runs your AI app receives what you paste. Swap the details for labels, and type the real detail when you paste.
  • A saved prompt can get worse replies than it once did, because apps, and the AI behind them, change over the months. Paste the prompt into a new chat, compare the reply with the line beside the prompt, and reword the prompt in your note.
  • A pile of saved prompts you never open stops helping, because you cannot find the one you want. Keep the few you reuse, begin the line beside each prompt with the job it does, and delete the other prompts.
  • A feature for saving prompts inside one AI app keeps your prompt in that app. The app can change or remove the feature, so keep a copy of the prompt in a note you own as well.

Why it matters

A prompt that worked is work you have already done once. Saved in a note you own, it pastes into any AI app again, so you do not do the work twice.

Related concepts

Outcomes

  • keeping-an-ask-1 (Keeping it running, waypoint 1): I can keep a prompt that worked as a note I own, with what it is for. Proves: after a prompt worked, they paste its words from a note outside the chat, with the line beside it on what it is for and what a good reply looked like (private details replaced by labels); you see the note's text in the chat.
  • keeping-an-ask-2 (Keeping it running, waypoint 1): I can reuse a saved prompt and tell when it has stopped working as well. Proves: they paste a saved prompt into a new chat, compare the reply with the line beside it, and say one thing they would change in the note (or that it needs nothing); or, on the fast track, they say what they would do when a saved prompt slips.

What success looks like

The second time they need a thing they have done before, they paste from a note instead of retyping. They have two or three saved prompts, each with a line on what it is for. When one slips they reword it rather than abandon it, and "how did I ask that last time?" comes up less.

Fast track

  • Explains: the words of a prompt that worked are plain text they keep. They keep the single prompt (not the first message) in a note they own, so it works in any app, with a line beside it on what it is for and what a good reply looks like. A person who names one app's save feature as the home, and cannot say what they would do if it went, is not yet there.
  • Ask: "You saved a prompt last month that gave you a great email reply. Today the reply is worse. What do you do?" Shows it: pastes the prompt into a new chat, compares the reply with what a good reply should look like, rewords the note and keeps the new version; does not blame themselves or drop the prompt without trying.
  • Ask: "Here is a chat: 'Write a reply to Dana about her late invoice 4471', then 'Shorter, and less formal', then 'Use first names, and mention we spoke on Tuesday'. Which words would you save, and what goes beside them?" Shows it: one clean prompt carrying the task and the changes together (short, informal, first names, the Tuesday call), the name and invoice number swapped for labels, and a line on what it is for and what a good reply looks like. The first message alone, or all three pasted as they stand, is not a pass.

Ways to teach it

Hints for the person's AI. Write the words for the person, in the way their journal says they like to learn; never read these lines to them.

  • Experiment: the Try it above. A second one, for a person who wants more (4 minutes): paste the same saved prompt into two different AI apps, or into one app a month apart, and write the one thing that differs into the line beside the prompt.
  • Challenge: for two weeks, each time a prompt works after a correction, ask your AI app for the single prompt and save it with the line beside the prompt. At the end, open the note, keep the prompts you reused and delete the rest.
  • Quiz (only if their journal allows quizzes): from the person's own use, three questions. The chat where a prompt finally worked and which words to save. A saved prompt naming a customer and what changes before it goes in a shared note. A saved prompt that stopped working and two things to try before deleting it.
  • Cards: four, on the words that worked being plain text, a note you own going to any app, the line beside the prompt saying what good looks like, and checking a saved prompt again when replies slip.
  • Explain: take the last prompt the person typed twice and show them what the note would have held. Ask them for the line beside the prompt rather than supplying it.
  • Game, "what goes in the note?": three short sample chats, each with a correction or two and a private detail. The person writes the single prompt and the labels for each.
  • Project (only for a person who has a Rucksack, which is coming soon): gather their two or three reused prompts into one plain file in the Rucksack, each with the line beside the prompt and labels.

Watch out

  • Say "a note you own", "your notes app or a document". Never "prompt library", "prompt engineering", "bookmark" or an app's feature name to the person on first teaching; an app's feature is a shortcut to mention afterwards, only if they ask or are in that app.
  • Not custom instructions: instructions ride in every chat, so putting a prompt for one job in them is the overcorrection (custom-instructions). A project's instructions sit between, applying to every chat inside that one project only. Not memory either: memory keeps facts about the person, not an exact text (app-memory). Not a skill: a prompt with steps and inputs, run by any app, is a later page; do not teach it here.
  • Say "often" and "check it", never that a saved prompt works forever. The same words can give a different answer after the app or the model changes (arXiv 2307.09009; Wharton, below). No study measures whether a saved prompt helps; the single-prompt step is a technique the person checks for themselves in Try it, so never call it a finding.
  • A pile of saved prompts is the overcorrection: a 200-prompt library, tagged and never opened; Nielsen Norman Group found Claude power users' context files went stale, were rarely edited and could not be found (9 October 2026). Help the person keep two or three.
  • App features differ and are changing (all checked 23 October 2026). Never recommend building on a feature named as retiring. ChatGPT: custom GPTs are being retired, Enterprise on 11 December 2026 confirmed, other plans "expected to follow", undated; Free, Go, Plus and Pro cannot create new ones; skills are on Business, Enterprise, Healthcare and Edu, personal plans UNCONFIRMED. Claude: skills on Free, Pro, Max, Team and Enterprise with code execution on; styles UNCONFIRMED. Gemini: Gems end in November 2026 for personal accounts (they move to skills automatically; Google, 30 September 2026), March 2027 for Workspace business and June 2027 for education. Copilot is the only app with a plain saved-prompt feature, and what a personal account sees is UNCONFIRMED.
  • Daypack: the note is in the person's own notes app or a document, and the person does the saving. After a prompt works, offer to write out the single prompt and the line beside it, ready to copy. The Daypack setup offers once to keep the look-back prompt in a note of their own with its first line above it (start.md); at the look-back, offer a Daypack person who kept it the second half of tip 3: run it again and check the reply against the line. Their own boundary line (their Never line) decides what stays out of a saved prompt. Nothing is recorded: no proven mark, no met mark. (checked 27 October 2026)
  • Rucksack (coming soon, so never something a reader can set up today): the note can be a plain file in the person's Rucksack, which you can read, so when a task matches a saved prompt you can say so in one line, once a chat. Write the file only with their yes, and the boundary line applies: no passwords or secrets in a saved prompt. (checked 27 October 2026)

Questions people ask

  • Q: Where do I save a prompt that worked? A: In a note you own, in your notes app or a document, with one line on what it is for. Some apps have a feature for this, but features change, so keep the words in a note too.
  • Q: Is a saved prompt the same as custom instructions? A: No, because instructions ride along in every chat, while a saved prompt is for one job and you paste it in when you need it.
  • Q: Why does my saved prompt work worse now? A: Apps and the AI behind them change, so the same words can give a different answer. Paste it into a new chat, compare the reply with the line beside it on what good looks like, and change the note.
  • Q: Can colleagues use my saved prompts? A: Yes, share the note or use your app's sharing if it has any. Take out client names and private figures first, since they travel with the prompt.

For your AI

Facts (checked 23 October 2026 unless a line says otherwise; fact check research/2026-10-23-essentials/batch-1-facts.md; research note research/2026-10-20-concepts/keeping-an-ask.md; plan product/learning-system/plans/keeping-an-ask.md). The page names no app feature on purpose: features change, a note does not.

Teach it like this: when a prompt has worked after a correction or two, offer once to write the single prompt that would have got there in one go, then to keep it in a note. When the prompt is lost ("it worked last week"), look back through their chat history for that chat if your app can, or rebuild the prompt together from what the good reply had. A habit for every chat belongs in their instructions, and a prompt for one job belongs in a note they own. Name an app's saving feature only if they ask, from the facts above, and say it may change.

An example of this taught well, in passing, after a customer reply took three tries: "That took a few goes, and the version that finally worked is worth keeping. Want me to write the one prompt that would have got it first time? Then you can paste it into a note, with a line on what it is for and what a good reply looks like, and use it next month in any app. Swap the customer's name for 'the customer' so it stays out of the note."

keeping-an-ask.md