Every team that uses AI ends up with the same problem: prompts scattered across chats, docs and personal notes, each one a slightly different copy of the original. Nobody knows which version is current. We built our own prompt library to get out of exactly this situation.
The fix is less about tools and more about a few shared habits: one place, clear names, an owner and a test date. Here is the setup we use, and why it builds on templates with variables rather than finished prompts.
One place, not five
If a prompt lives in a chat thread, it is already lost. Pick one shared place, whether that is a library, a wiki page or a repository, and agree that anything not there doesn’t officially exist.
Names that sort themselves
A consistent title format makes a long list scannable. We use verb + object + context:
Write product description – e-commerce Review pull request – security focus Summarise support ticket – for escalation Generate blog header – flat illustration
Organise by project and task, not by person
Folders named after people stop working the day someone changes roles. Group prompts by project (which client or product it belongs to) and task type (what it does). That way anyone can find a prompt without knowing who wrote it.
Every prompt has an owner and a test date
- Owner: the person who answers questions and approves changes.
- Status: untested, tested or deprecated.
- Tested with: the model and date of the last real test.
A prompt marked deprecated is still useful as history, but nobody copies it by accident.
Shared pieces live once
Personas, project context and brand rules are used in many prompts. Store them once as reusable snippets and include them where needed. When something changes, you update it in one place, not in thirty prompts.
A shared library is only as good as its worst-maintained prompt.
Operations manager
A monthly ten-minute review
- Sort by least recently updated.
- Re-test the five most-copied prompts.
- Deprecate anything nobody has used in three months.
- Merge duplicates and keep the better-tested one.