Building a prompt library your team will still use in six months

Nearly every team we train leaves with a prompt library. Some are still in daily use a year later, and most are not. The difference is not the quality of the prompts. It is whether the library was built to be used or built to be delivered.

Organise by task, not by tool

A library sorted into "ChatGPT prompts" and "Copilot prompts" forces the reader to know the answer before they can look it up. Sort by the job instead: replying to a complaint, writing a job ad, summarising a call. People search for what they are trying to do.

Keep the example attached

A prompt without an example of its output is an instruction people have to trust. With a real before-and-after attached, someone can tell in five seconds whether it is the thing they need. Store the example next to the prompt, from your own work, not a generic sample.

Give it an owner and a review date

Models change, and a prompt that was carefully tuned last spring may now be over-engineered or plainly worse than a simple request. Without a named owner, nobody notices. Put a person and a quarterly review against it, or accept that it will quietly rot.

• ✳Sort by task, so people can find things without already knowing the answer
• ✳Attach a real example of the output to every entry
• ✳Name an owner and a review date
• ✳Delete entries that stop earning their place