Why shared prompt docs stop being used
Most team libraries start as a shared doc of prompts that worked once. Then they grow by copy and paste. Someone changes a line for their case and saves it as a new prompt, and a customer’s name from a real case stays in. The first prompt anyone wrote for a task stays the one everybody uses, because nobody compares it with the others. After a few months nobody knows which of four similar prompts to trust, so people go back to typing their own.
A library people use is the opposite: fewer prompts, each better than what people would type themselves and easy to find. Every prompt in it falls somewhere on two lines, how well it is written and how often the team reaches for it, and that decides what to do with it.
- Board update: How often the team uses it 2/10, How well it is written 9/10
- Support reply: How often the team uses it 8/10, How well it is written 8/10
- Old launch email: How often the team uses it 2/10, How well it is written 2/10
- Client report: How often the team uses it 8/10, How well it is written 2/10
- Job ad: How often the team uses it 6/10, How well it is written 6/10
Illustrative: example prompts placed by judgment. The Library Audit scores how well each prompt is written; how often each one is used, your team can tell you.
1. Decide what an entry holds
An entry needs more than the prompt. Someone opening it for the first time should see within seconds whether it fits their task and what to paste in:
Name: Reply to a customer whose order
arrived damaged
Use it when: a customer reports damage, with or
without photos
You paste: the customer's email and the order
details
Owner: Support lead, with a named backup
Score: 9 of 10 on the ten checks
Reviewed: November 4, 2026
Prompt: Write a reply to a customer whose
order arrived damaged...
Add a sample of a good answer when the format matters. It shows people what to expect, and it gives the owner something to compare new answers with. Keep the library where the team already works, such as a wiki page or a shared drive folder, and link it from the places where the work happens.
2. Name each prompt for the job
Name prompts the way people will search for them. Start with a verb and the thing it makes, and add who it is for when that matters: “Reply to a late-delivery complaint” is easy to find, and “Support prompt v3 final” is not. Group prompts by the job, such as support replies or hiring, not by which AI they were written for.
Keep version numbers and dates out of the name. They belong in the entry, where the owner updates them; in a name, they multiply into copies.
3. Give each prompt one owner
Each prompt needs one named person, with a backup, usually whoever lives with its results: the support lead owns the reply prompts, and the finance manager owns the invoice reminders. The owner approves every change to the live version, keeps a few real test inputs, answers questions about it and decides when it retires. A prompt nobody owns is the first to go stale. The guide to rolling out prompts to a team covers version numbers and testing a change before everyone gets it.
4. Keep customer details out
A prompt that worked on a real case often still carries it: a customer’s name, an email address, an order number, an account balance. Everyone who copies the prompt copies those details too, into whichever AI tool they use. Replace each one with a named blank, such as {{customer_message}} or {{order_details}}, and say in the entry what goes in it.
The same goes for anything that should never leave the company, such as keys, passwords, internal codes and salary figures. When people later paste real material into a filled-in prompt, Safe Paste swaps the private details for placeholders before it reaches the AI, and puts them back in the reply.
Before
2/10 Needs work
- Has: What to make
- Missing: Who it's for
- Has: Background facts
- Missing: Length
- Missing: Format
- Missing: What to avoid
- Missing: What good looks like
- Missing: An example
- Missing: If unsure
- Missing: Layout
After, with what was added marked
9/10 Strong
- Has: What to make
- Has: Who it's for
- Has: Background facts
- Has: Length
- Has: Format
- Has: What to avoid
- Has: What good looks like
- Missing: An example
- Has: If unsure
- Has: Layout
Illustrative; the customer and the order are made up. The first version is one real case saved as a prompt, with the customer’s details still in it. The second turns each detail into a named blank and adds the policy, the length, the format and what to do when a fact is missing, so it works for the next case too. The marked words are what it adds.
5. Score prompts before they go in
Set a bar a prompt must clear to join the library, and check it before people rely on it, not after. A workable bar is a score of at least 8 out of 10 on the ten checks in the Prompt Grader, which is the lowest score the prompt repository allows for its own prompts, plus good answers on a few real inputs.
A score tells you whether the prompt gives the model what it needs: the task, the reader, the facts, the format and what to do when unsure. It cannot tell you whether the answers are right for your work, which is why the real inputs matter.
6. Retire near-copies
Near-copies are prompts that share most of their wording and do the same job, usually because someone changed a line and saved a new one. Keep the strongest and fold any useful line from the others into it. Then replace each copy with a note pointing to the one you kept, for a month, so old bookmarks still lead somewhere.
The Library Audit groups near-copies however they are named, and suggests which one to keep.
Delete the weak one today
Not quite. People paste their own copy back from an old chat, and you lose whatever made it the favorite, which is often a shorter set of blanks.
Keep both, since people like a choice
Not quite. Two prompts for one job is how libraries fill with copies, and nobody can tell which one to trust.
Fold what people like about the popular one into the strong one, and point the old name at it
Yes. The strong prompt gains what made the other popular, and nobody’s bookmark breaks.
Ask the team to vote
Not quite. Popularity measures habit, not quality. Compare the answers each one gives on the same real inputs.
7. Find the gaps
A library that covers the wrong tasks goes unused however good it is. Ask the team what they used AI for last week, or what they still do by hand every week, and compare the answers with the library. Tasks that come up often with no prompt behind them are the gaps to fill first.
The Library Audit also compares your prompts with the 669 expert prompts in the repository and lists the ones in your areas that the library does not have yet. Start a missing prompt from one of those and fill in what stays the same for your team.
8. Review it every quarter
Once a quarter, whoever keeps the library spends an hour on it:
- Run the whole library through the Library Audit. Paste the prompts, or use the ones saved in My Prompts. You get a health score from 0 to 100, how the scores spread, the near-copies, any private details and the prompts to fix first. The prompts are read on thinkhard.si and then forgotten.
- Fix or retire the prompts it lists first.
- Merge the near-copies into the one to keep.
- Check every prompt still has an owner who is still on the team.
- Run the most used prompts on a few real inputs, because the model underneath may have changed since the last review.
Write the date and the health score at the top of the library. A number that rises from one quarter to the next shows the team the library is looked after, which is part of what makes people trust it.
Key takeaways
- A library people use is short and trusted: fewer prompts, each better than what people would type themselves.
- Give every entry a plain name, a line on when to use it, what to paste in and one owner.
- Replace customer details with named blanks before a prompt goes in.
- Set a bar to get in, such as 8 out of 10 on the ten checks and good answers on real inputs.
- Retire near-copies into the strongest version, and fill gaps from the tasks the team does every week.
- Review the whole library every quarter and write down its health score.
Questions
What is a prompt library?
A shared collection of prompts a team reuses for its regular work, such as support replies or job ads, each with a name, a note on when to use it, an owner and blanks for the details that change. It can live in a wiki page or a shared drive; what matters is that everyone copies from the same place.
Where should a team keep its prompt library?
Wherever the team already works and can find it in seconds, such as a wiki page or a shared drive folder, linked from the places where the work happens. Everyone reads it, and owners edit it. My Prompts keeps your own versions in your browser, which suits drafting, but the team’s copy belongs in a shared space.
How many prompts should a team library have?
There is no right number. Start with the tasks the team does every week, which is often a short list, and add a prompt when the same gap shows up more than once. A short library people trust is easier to keep current than a long one they have to sort through.
How do I stop people keeping their own copies?
Make the shared version better than theirs and easy to find. Give each prompt an owner who takes suggestions and folds good changes back in quickly, and when you retire a copy, leave a note pointing to its replacement.
Is it safe to keep prompts in a shared library?
Yes, as long as the prompts hold no customer details or passwords. Replace real names and account numbers with blanks before a prompt goes in, and check the whole library each quarter. The Library Audit flags email addresses, phone numbers, card and account numbers, keys and passwords.
Are there free prompt libraries to start from?
Yes. The thinkhard.si prompt repository has 669 prompts for work, grouped by job, each scored on the ten checks. Copy the ones that match your team’s tasks and fill in what never changes for your team, leaving the rest as blanks.
Prompts to use with this
- Improve a prompt that gives inconsistent resultsYou need a prompt you use often rewritten so it gives reliable, well-formatted answers.
- Build reusable templates for emails you send oftenYou write the same kinds of email every week and want saved templates with clear blanks, in your own voice.
- Turn a messy process into a standard operating procedureA process lives in one person's head or a chat thread, and you need a checklist anyone on the team can follow.
- Draft an AI acceptable-use policy for staffYou need clear rules on which AI tools staff can use, for what, and with what data.