Keep your best prompts in one place and stop starting from scratch

The prompt that finally gave you a usable meeting follow-up is somewhere in last week’s chat history, between a spreadsheet question and a lasagna recipe. Next time you need it, you’ll write a new one from memory, and it will be a little worse than the one you lost.

A prompt that worked has already been through the expensive part. You found the wording, saw what went wrong the first time, added the rule that stopped it and checked the answer. Retyping it from memory throws that work away. The fix is a little old-fashioned: one document with a card for each prompt worth keeping, two clicks from wherever you open your assistant.

By the end of today, in about an hour, you’ll have that document with five tested cards in it, and a one-line check you can drop into any chat in a couple of keystrokes.

One card per prompt

A card holds the prompt and everything you’d otherwise have to remember about it. Copy this format:

NAME:        [a short name you'd search for]
USE IT WHEN: [the situation, in one line]
FROM:        [the post or tutorial it came from, if any]

PROMPT:
[the prompt, with every part that changes in [SQUARE BRACKETS]]

CHECK IN THE ANSWER:
- [what this prompt tends to get wrong, and how you catch it]

TESTED: [date], in [assistant], on [the made-up or public material you used]
TIMED:  [minutes from opening the library to a checked answer, and the date]

Two habits make the cards work. Write every part that changes in capitals inside square brackets, such as [READER], and search the prompt for “[” before you send, so no blank slips through. And write the check while you still remember what went wrong; it’s the line you won’t be able to reconstruct in three months.

For example, here is a card for a small prompt that earns its place most weeks. The TESTED and TIMED lines are yours to fill in.

NAME:        Shorten a message
USE IT WHEN: a draft email or update runs long and I need it cut without losing a request

PROMPT:
Shorten the message below to under [NUMBER] words.
- Keep every request, name, date and number.
- Put what I need from the reader in the first two lines.
- Add nothing new.
- After the shorter version, list everything you cut.

Message:
[PASTE THE MESSAGE]

CHECK IN THE ANSWER:
- Every date and request in the original is still there.
- The list of cuts matches what's missing. Put back anything the reader needs.

TESTED: [today's date], in [your assistant], on a made-up three-paragraph update
TIMED:  [the first time you use it on a real message]

The last rule in the prompt does the most work. An assistant asked to shorten can drop a deadline as easily as a pleasantry, and the list of cuts tells you which one it dropped, so you don’t have to reread both versions side by side.

Where to find your first five

Start with your own. Scroll or search your chat history for the requests you make every week; the verbs give them away, such as shorten, rewrite, reply, summarize or “turn this into a table”. Any prompt you’ve rewritten from memory twice has earned a card.

Then borrow from this series. Each post below has a prompt with rules against the mistakes assistants usually make, and checks you can copy onto the card.

Email

• A hard message, rewritten from your blunt draft: The email you’ve been putting off.
• A reply in your team’s voice, with markers wherever a person has to decide: A style card for replies.

Meetings and weeks

• A follow-up from notes marked D, A and Q: Meeting follow-up within the hour.
• A status report from a log you keep all week: A running log that turns into your weekly status report.

Numbers

• A spreadsheet formula from a plain description: Describe the formula you need.
• Commentary on figures the spreadsheet already calculated: Let the spreadsheet do the math.
• Themes for a pile of open comments, which you count yourself: Comments sorted into themes.

Documents

• An interview before the assistant drafts anything long: Why your AI drafts sound generic.
• One headline per slide, before any design: Write the slide headlines first.
• A translation that follows your own terms: Better AI translations start with a short glossary.

Two clicks or fewer

Count the clicks between a blank chat and the prompt on your clipboard. A library more than two clicks away is the first thing you skip in a busy week, and then you’re back to retyping.

• A document. One file called “Prompt library”, with each card’s NAME as a heading so the outline or navigation pane lists every card. Bookmark it or pin it wherever you pin files. It’s also the easiest version to share.
• A notes app. One note per card in a single folder, handy if you also use an assistant on your phone.
• Text replacement on a Mac or iPhone, for one-liners. On a Mac, choose Apple menu > System Settings > Keyboard > Text Replacements. With iCloud Drive on and the same Apple Account on each device, they also reach your iPhone and iPad, where the list is in Settings > General > Keyboard > Text Replacement.
• Clipboard history on Windows, for one-liners. Press Windows logo key + V, select Turn on the first time, and pin the items you reuse. Pinned items stay after a restart; the rest of the history is cleared.

A good first one-liner is a check that several prompts in this series build in, in one form or another. Save it as a text replacement with a code you’d never type by accident, such as ;src, or pin it in your clipboard history:

After your answer, list every fact, name, number or date that doesn't come from what I gave you. Write "None" if there are none.

Try the code once in the window where you actually write prompts. If nothing happens there, a text expander app does the same job on either system. (Menus checked in October 2026.)

Some assistants have their own place for saved prompts or projects. Use it if you like; your own library works with any assistant and keeps the checks next to each prompt when you switch.

Your role and style go in the settings

If many of your prompts open with the same paragraph about who you are, what your team does and how you like answers written, move that paragraph to your assistant’s standing instructions. Longer material, such as your style card or glossary, can sit in a project as reference files. Set up your AI assistant once walks through both. The library keeps the jobs themselves, such as this week’s follow-up or the monthly report.

Sharing it with a team

A shared library saves everyone from rediscovering the same wording, and it goes stale quietly unless somebody looks after it.

• One owner adds, edits and retires cards; everyone else suggests changes in a comment. Add an OWNER line next to TESTED, so people know whom to ask.
• A “Retired” section at the bottom, with one line on why each card left (“a template does this better”), so nobody adds it again next spring.
• The test material stays in the library, so anyone can rerun a card. If a prompt handles details about clients or colleagues, agree on what may be pasted; Before you upload a client document lists the questions.

Retest every three months

Assistants change their models and defaults, and a prompt that held its word limit in March may ignore it by September. Every three months, and on the day you switch assistants, run each card once on its test material, compare the answer with the checks, fix what slipped and change the date.

Two signs mean a card should change shape. If you edit the prompt before every use, add brackets for the parts that change, or split the card in two. If you rewrite the answer into the same words every time, save those words as a template, which needs checking only once (Five ways out of a repetitive task shows when a template wins). A card you haven’t used since the last retest goes to Retired.

Try it today

Set aside an hour.

1. Create the document or note and call it “Prompt library”. Bookmark or pin it, then count the clicks from a blank chat to a prompt inside it. More than two means it needs a better spot.
2. Paste the card format at the top.
3. Choose five prompts: any you’ve rewritten from memory at least twice, plus the ones above that match this week’s work. Make a card for each.
4. Test each card once. Use made-up or public material, or real work if you’ve checked what your assistant does with it. Fill in TESTED only when the answer passes the card’s checks.
5. Save the ;src check as a text replacement or a pinned clipboard item, try it once in a chat, and put the three-month retest in your calendar.

You’re done when the library holds five tested cards, each with today’s date, two clicks from a blank chat. The next time you use one for real, time it from opening the library to a checked answer and write the minutes on its TIMED line.

A month from now

If you haven’t yet, do How to find your most repetitive work this week: it ranks your recurring tasks by minutes per month and has you time the first one you fix. Give its review prompt a card, because you’ll use it again. A month from today, repeat the audit, time the fixed task again and see what now tops the list. Run that task past the five exits, and if an AI draft is the right one, its prompt already has a place waiting in your library.