Blog / Build with AI /

Prompt libraries: saving the prompts that work

You wrote a prompt last month that produced exactly the output you wanted. Clean, in the right format, no waffle. You used it, felt pleased and moved on.

Today you need it again. It’s somewhere in a chat history with four hundred other chats. You scroll, search, give up and rewrite it from memory. The new version is worse.

That small frustration has a simple cure: a prompt library, a place where the prompts that work are saved on purpose, with enough notes that future you can use them.

It sounds like admin. It’s closer to a toolbox. A carpenter doesn’t re-forge a chisel for every job.

Why bother

Consistency. The same prompt gives you results of similar quality and shape every time. That matters for anything repeated: product descriptions, reports, data cleaning, code reviews.

Speed. Starting from a tested prompt takes a minute. Starting from scratch takes ten, and the result is a gamble.

Learning. When you review old prompts, you see what you’ve learned. You spot the habits that work: examples, constraints, formats.

Sharing. A team with shared prompts doesn’t depend on whoever happens to be good at prompting.

Switching. If you move between ChatGPT, Claude, Gemini or something new, your library comes with you. The tool changes; the know-how stays.

Where to keep it

The best place is the one you’ll actually open. Options, from simplest:

  • A notes app (Apple Notes, Google Keep, Notion, Obsidian). Easy, searchable, works on phone and computer.
  • A document in Google Docs or Word, with headings by category.
  • A spreadsheet. Good once you have dozens of prompts, since you can filter and sort by task, tool or rating.
  • A plain text or Markdown file in a folder you sync. Portable and tool-independent.
  • Features inside AI tools (saved instructions, custom assistants, project instructions). Handy, but they live inside one product, so keep a copy elsewhere.

My advice: start in a spreadsheet or a notes app, and don’t spend a week choosing. The system should take five minutes to set up.

What to save with each prompt

A prompt on its own is half the value. Save these fields with it:

Field Why it’s useful
Title A name you’d search for: “Clean product titles”
The prompt The exact text, with placeholders in brackets
Tool and model Prompts behave differently across models
Example input A sample of what to paste in
Example output What “good” looks like
Notes What to watch for, what it gets wrong
Date and version So you know if it’s current
Rating A quick 1 to 5

The example input and output are the most valuable. They let you check, in seconds, whether the prompt still behaves as expected after a model update.

A template to copy

Title: Rewrite a product title for search

Prompt: You are helping an online shop improve product titles. Rewrite the title below to include the brand, product type and the single most important attribute. Maximum 60 characters. Don’t add claims that aren’t in the original. Return only the new title.

Title: [PASTE TITLE HERE]

Tool: Any large chat model, tested Feb 2025

Example in: “BLK wireless mouse 2.4g usb nano rec.”

Example out: “Wireless USB Mouse, Black, 2.4GHz”

Notes: Occasionally invents a brand name when the original has none. Check results.

Notice the square brackets for the part that changes. A good prompt is a template, with fixed instructions and a clear slot for the variable.

How to organise it

Don’t over-engineer. Three levels are plenty.

By job:
– Writing and editing
– Data and spreadsheets
– Code and debugging
– Research and summaries
– Customer replies
– Planning and brainstorming

By status:
– Tested and trusted
– Draft or experimental
– Retired (keep, but mark)

By tool: only if you find that prompts differ a lot between assistants.

Tags help more than folders. A prompt can be both “data” and “spreadsheets”.

What makes a prompt worth saving

Not every prompt deserves a place. Save one when:

  • You’ve used it more than once, or you know you will.
  • It took effort to get right.
  • The result is easy to check.
  • It encodes a lesson, like a rule you had to add after the first attempt went wrong.

Don’t save one-off questions. A library of three hundred random prompts is a junk drawer.

Things that make prompts reusable

  • Use placeholders. [TOPIC], [AUDIENCE], [PASTE TEXT HERE].
  • State the format. “A table with three columns”, “one line per item”.
  • Give an example. One good example does more than a paragraph of explanation. See giving AI examples: the fastest way to better output.
  • State the constraints. Length, tone, what not to do.
  • Say who the output is for.
  • Keep the personal data out. A library often gets shared. Use placeholders instead of real names, emails or client details.

Keeping it alive

A library that isn’t maintained decays. Models change and old prompts that worked beautifully start producing odd results.

  1. Re-test your top ten prompts whenever a major model update arrives, or every few months. Use the saved example input and compare with the saved example output.
  2. Version your edits. Change the date and write one line about what changed and why.
  3. Retire the dead ones. Mark them, don’t delete, in case they teach you something.
  4. Add as you go. When you finally get a prompt right, save it that minute. Not later.
  5. Review once a quarter. Ten minutes. What did you use? What did you rewrite? What’s missing?

For teams

If several people use AI, a shared library is one of the cheapest productivity wins available.

  • Put it somewhere everyone can read, with one or two people allowed to edit.
  • Add an “owner” and “last tested” date to each prompt.
  • Agree on a naming style.
  • Include a “do not paste” list: customer data, passwords, anything covered by an agreement with a client.

If your tools work with an API, your prompts become part of your product. In that case, they belong in version control or a configuration file, not in a chat history. See version numbers and changelogs for small tools.

Prompts for building things

If you build tools with AI, these are the prompts I’d save first:

Add them to the library, then improve them each time one lets you down.

Common mistakes

  • Saving the prompt but not the result. You can’t tell later whether it worked.
  • No placeholders, so every reuse needs editing.
  • Too many prompts, no ratings. You stop looking.
  • Locked into one tool. Keep a copy outside it.
  • Sharing a library with private data in it.

Key takeaways

  • A prompt library is a place to keep prompts that worked, with notes on what good looks like.
  • Save the prompt, an example input and output, the tool and a date.
  • Keep it simple, in a notes app or spreadsheet, and use placeholders.
  • Re-test your best prompts every few months, because models change.

What is the one prompt you rewrite from memory every week, and why isn’t it saved yet?