The prompt library is where good questions are kept. One person works out how to ask something properly, and everybody else gets that version instead of inventing their own.
Why it exists
The difference between a useful AI answer and a useless one is usually the question.
"Summarise this" and "summarise this in three bullets, focused on what we committed to and by when" produce answers of completely different value from the same model. The second took somebody two minutes to work out.
Without somewhere to keep it, that two minutes is spent again by every person who needs the same thing, and most of them settle for the worse version.
Placeholders
A saved prompt can leave gaps to fill in when it is used.
A quarterly review prompt with the quarter and the team left blank is one prompt used all year by everybody, rather than nine near-identical prompts that gradually drift apart.
Leave a gap wherever the answer changes. The name, the period, the product, the customer. Everything else stays fixed, which is the part that makes it reusable.
Organising it
Prompts are filed by category and by the area they belong to, which is how they stay findable at more than about twenty.
File by who needs it, not by what it does. Somebody in Finance looks under Finance. They do not look under summarisation.
Name it as the job. "Draft a payment chase" is found. "Prompt 3 - AR" is not.
Using one
Copy it into the assistant, in the cockpit, in the command bar, or in comparing models when testing.
That is deliberately simple. A prompt is text, and text works everywhere the assistant does.
Prompts and knowledge do different jobs
A distinction worth being precise about, because getting it wrong wastes a lot of effort.
A prompt shapes the answer. It says what to produce, how long, in what form, and what to leave out.
Knowledge supplies the facts. It is what the assistant knows about your organisation.
The failure looks like this: somebody gets a generic answer, assumes the question was badly phrased, and spends an afternoon rewriting the prompt. The prompt was fine. The assistant had nothing of yours to draw on, so the best possible phrasing still produced a well-shaped generic answer.
Test which one you have by asking something only your organisation could answer. If it cannot, no prompt will save it, and Teaching the assistant is the work.
Variables, in practice
The gap-filling is worth using properly rather than treating as optional.
Name the gaps clearly. A gap called "quarter" is filled correctly. One called "input" is filled with anything.
Keep the number small. Two or three gaps is a template. Eight is a form, and people abandon it and type their own question.
What makes a prompt worth saving
Four things, and a prompt missing all four is a note rather than a tool.
It says what output is wanted. Length, shape, format. "Three bullets" beats "briefly" because briefly means four different things to four people.
It says what to focus on. The same document summarised for a lawyer and for a salesperson should produce different summaries, and only the prompt can say which.
It says what to leave out. Often the most effective line. "Do not restate the background" removes a paragraph nobody reads.
It has been used more than once. A prompt written for a single occasion is not a library entry, and a library full of those is one nobody browses.
Keeping it useful
A prompt library rots faster than most people expect, because a prompt that has stopped working still looks fine.
Delete rather than keep. Twelve good prompts get used. Ninety, of which twenty work, get abandoned as a set, because the cost of finding the good one exceeds the cost of writing your own.
Re-test after changing the model. Prompts are tuned to a model more than people realise, and one that produced tight three-bullet summaries can start producing essays after a switch. That is a real thing to check after reading Choosing your model.
Review when the vocabulary changes. A prompt naming a stage or a product that has been renamed produces confidently wrong output rather than an error.
Keep the good ones at the top
Order matters more than volume in a library people browse rather than search. The three most used should be the three most visible.
Sharing
A prompt is written once and used by everybody, which is the entire value.
Save it for the team, not for yourself. A private prompt helps one person and leaves the same problem unsolved for the other nine.
Say who it is for in the name. "Draft a payment chase (Finance)" tells somebody browsing whether to read further.
Where to start
Look at what people actually ask the assistant, which is visible in usage, and save the good version of the three most common.
That is a better starting point than designing a library, because it is grounded in what your organisation genuinely needs rather than what somebody imagined it would.
Two habits
Save it the moment it works. The good phrasing is arrived at by iteration, and nobody reconstructs it later. If you got a genuinely good answer, the prompt that produced it is worth thirty seconds.
Improve in place rather than adding. A second prompt that is nearly the same as the first is how a library becomes unusable. Edit the original.
Two people, one prompt
If two people keep asking the same thing differently, that is the signal to save one. The library grows best from observed repetition rather than from planning.
What it is not
It is not knowledge. A prompt asks well; it does not tell the AI anything about your organisation. That is Teaching the assistant, and a beautifully written prompt against an empty knowledge base still produces generic answers.
It is not a script. Prompts are starting points people adapt. A library treated as mandatory phrasing gets ignored.
Did this answer your question?
No, ask a person