Why good prompts get lost
A prompt that works usually gets written once, in the middle of a task, inside a chat. A week later it sits in a long chat history, mixed in with everything else you asked that day. You remember it worked, but not the exact wording, so you write a weaker version from memory.
The fix isn't a cleverer folder structure. You need three habits: save a prompt the moment it proves useful, store it somewhere you can search, and make it quick to paste back into whichever AI tool you're using.
Where to store AI prompts
There's no single right place. The best home depends on how many prompts you have and how many tools you use them in.
| Where | Works well for | Breaks down when |
|---|---|---|
| Built-in assistant features (custom instructions, projects, GPTs, Gems) | Standing instructions for one assistant | You use more than one assistant, or need dozens of task prompts |
| Notes app or a single document | Getting started with a handful of prompts | The page gets long, and finding and copying the right prompt is slow |
| Spreadsheet | Cataloging prompts with columns for model and use case | Prompts are multi-paragraph, or you want to copy one in a hurry |
| Markdown files in Git | Developers who want diffs and reviews for prompts used in code | You need prompts on your phone or inside a chat window |
| Text expander or snippet app | A few prompts you type every day | You want to browse, search, or fill in more than one blank |
| Dedicated prompt manager | A growing library you reuse across tools and devices | You only ever reuse two or three prompts |
Built-in features are useful for standing instructions, but each one only works inside its own assistant. Our guide to saving prompts in ChatGPT, Claude, Gemini, and Copilot covers what each one can and can't do. Whatever you choose, keep one master copy of each prompt in a place you control and can export.
Name prompts by the job they do
You'll find prompts by searching, so name them the way you'd search. Start with the action, then what it acts on, then anything that makes it different from similar prompts.
- Vague: “Email prompt 3”, “Good one”, “Marketing”
- Searchable: “Summarize: sales call notes into next steps”, “Rewrite: email reply, firmer tone”, “Image: product photo on a white background”
Pick one pattern and stick to it. A consistent pattern matters more than which one you pick.
Use a few collections and many tags
Folders (collections) and tags do different jobs. Mixing them up is the most common reason a library becomes hard to use.
- Collections answer “what area of my work is this for?” Writing, Code, Research, Images, or one per client. Keep them shallow, at most one level of nesting, so each prompt has an obvious home.
- Tags answer “what is true about this prompt?” The model it's tuned for, the output type, and its status. A prompt lives in one collection but can have several tags.
A small set of status tags goes a long way: #best for your proven prompts, #draft for ones still being tested, and #template for prompts with blanks to fill in.
Turn repeat prompts into templates
If you've edited the same prompt for different inputs three times, turn it into a template. Replace the parts that change with named placeholders, so you fill in only the blanks next time instead of rewriting the prompt:
Write a {{tone}} cold email to a {{role}} at {{company}}.
Open with their likely problem: {{pain_point}}.
Keep it under 120 words and end with one question.Use descriptive placeholder names. {{audience}} tells you what to type; {{x}} doesn't. One template with four blanks replaces a dozen near-duplicate prompts that slowly drift apart.
Keep the result next to the prompt
A prompt on its own doesn't tell you what it produces. Save an example output with it: a sample answer, a code snippet, or the image it generated. This matters most for image and video prompts, where small wording changes have big effects. Note the model you used too, because the same prompt can behave differently in another model.
Keep the version that worked
Prompts get tweaked constantly, and the tweaks don't always help. Before you overwrite a prompt that works, keep the previous version, either as a saved history or as a dated copy. Add one line about what you changed and why. When a change makes results worse, you can go straight back.
Make prompts reachable where you type
A library you have to switch apps to open won't get used. The best setups put prompts one shortcut away from where you type: a browser side panel next to the chat, a global hotkey on your computer, or a keyboard on your phone. Each time you have to switch windows and copy, you're more likely to retype the prompt from memory.
Prune once a month
Ten minutes a month keeps a library useful:
- Merge near-duplicates into one template.
- Delete one-off prompts you never reused.
- Tag the prompts you actually reached for as
#best. - Retest key prompts after a major model update.
A 30-minute setup
- Pull the prompts you reuse from your chat histories, notes, and docs into one place.
- Rename each one with the action-object-detail pattern.
- Create three to six collections for the main areas of your work.
- Add model, output-type, and status tags.
- Turn your three most-repeated prompts into templates.
- Set up the shortcut you'll actually use: a browser extension, a hotkey, or a phone keyboard.
How this looks in PrompX
PrompX is a prompt manager built around this system. Prompts live in collections with tags, and you can search and filter by category, model, or output type. Templates use {{variables}} that you fill in before copying, and each prompt can keep its example outputs: text, code, or images. Pro adds version history with compare and restore.
Your library is reachable where you work: in a Chrome side panel next to ChatGPT, Claude, or Gemini, with ⌘⇧Space on Mac, and from the PrompX keyboard on iPhone. You can export everything as JSON from the web app at any time.