Building your first prompt library
How to turn the prompts that work into a shared team asset.
Most prompt libraries die within a month, and they die for the same reason. The prompt was never the thing that made it work.
The prompt you saved was maybe a tenth of it. The rest was the specific thing you pasted underneath - the brief, the transcript, the three examples of copy you liked, the customer list. Save the instruction on its own and you have kept the label off the jar and thrown the contents away.
What actually goes wrong
You know the shape of it. Somebody enthusiastic builds a Notion page. Sixty entries by the end of the week, neatly categorised by team. By week three, four of them are in use. By month two, nobody opens the page, and the person who built it has quietly gone back to keeping their good prompts in a personal note.
Four causes, and only one of them is about prompts.
The context is missing, as above. There is no owner, so nothing gets updated when it drifts. There is no versioning, so when output quality drops nobody can tell you which version used to be good. And it is organised by team or by tool rather than by task, which means finding the right entry is harder than rewriting the prompt from memory.
Save recipes, not prompts
The reusable unit is bigger than a prompt. Call it a recipe. Six parts:
The task, in one line, written as the job to be done rather than the tool being used. The instruction itself. The context it needs, and crucially where that context comes from - which document, which export, which field. An example of good output. An example of a failure, and what caused it. And a named owner.
The example of good output is the highest-value part and it is the part everybody skips. Models follow examples considerably better than they follow adjectives. One paragraph of copy you actually approved does more work than four sentences describing your tone of voice.
The failure example is the second most valuable and the most unusual. It is also the only part that tells the next person where the edges are.
[illustrative] An edtech marketing team, roughly 2024, built a 60-entry library in a week. Four entries were still in use by month two. They rebuilt it as six recipes with real example outputs attached, and all six were still in use six months later. Same team, same tools, a tenth of the entries.
Harvest, do not design
Do not sit down to build a library. Go and find the one you already have.
Ask your team which prompts they have re-used more than twice. There will be three to six of them, they will be unglamorous, and they will mostly be about reformatting or summarising rather than anything creative. Those are your first entries. Write them up properly, with the context and the examples, and stop there.
Five entries people trust beat fifty nobody has tested. Everything in a good library has been used in anger by someone who needed it to work.
And keep it where the work happens. A library living in a tool nobody has open is a filing cabinet in a basement. (I have built two of those. Both immaculate. Both abandoned.)
The bit that will age
Here is something to plan for rather than be surprised by. A lot of prompt craft is a workaround for model limitations, and those limitations keep getting fixed. The elaborate scaffolding - the role-play preamble, the step-by-step coaxing, the formatting tricks - has been getting less necessary every year, and that trend is not going to reverse.
What survives model upgrades is your context and your examples. Your approved copy, your category definitions, your real briefs, your worked cases. Those are assets. The clever phrasing is a temporary patch, so do not over-invest in it and do not let anyone tell you it is a career.
Where I would push back
For a small team this is overhead. If three people are doing the work and they sit near each other, they can just ask, and a formal library is process for the sake of process. I would not bother below about five regular users.
There is also a genuine risk that a well-maintained library entrenches yesterday’s method. It quietly becomes the approved way to do something, six months after a better way arrived, and the better the library is the harder it is to argue with. Put a review date on it. Somebody should be allowed to delete things.
What to do on Monday
Find the three prompts already being re-used in your team. Write each one up with the context it needs and one real example of output you approved.
Three recipes, one owner, one place to find them. That is a prompt library. The other fifty-seven entries were never the point.
Editing notes
- No dates, vendor names or statistics in this piece, so nothing needs verifying. I deliberately did not name specific tools for storing a library, because that recommendation ages badly and invites “why not X” in the comments. Easy to add if you want it.
[illustrative]on the edtech 60-entries-to-6 example. The before-and-after shape is what makes the point, so a real version with your own numbers would be worth having.- “The bit that will age” is the strongest and most contrarian section. It also argues against the long-term value of the thing the post is teaching, which is a risk. I think that tension makes it better, but it is your call.
- The parenthetical about building two abandoned libraries is the piece’s one aside. Cut it if it is not true of you.