Short answer
To build a prompt library, take prompts your team has used successfully more than once and turn each into a template with clear variables, a named owner, the tool it was tested in, a test input with a known good output, notes on how it fails, data rules and a review date. Store entries in one shared place, keep the library small, and re-test entries whenever your AI tool or model changes.
- Only add prompts that have worked at least twice on real work.
- Every entry needs an owner, a test input and a review date.
- Record known failure modes next to the prompt, not in someone’s memory.
- Re-run test inputs after tool or model updates; outputs drift.
- A small team rarely needs more than 15 to 20 entries.
Once colleagues start asking “what was that prompt you used?”, a team is ready for a prompt library. Team leads, practice managers and internal AI champions will finish this lesson with three working prompts turned into entries that anyone on the team can pick up cold and get the same quality. Allow 40 minutes, your AI chat history, and one colleague willing to test an entry.
Why prompts need a library
A good prompt in one person’s chat history helps one person. Shared without notes, it gets edited, stripped of its rules and pasted with the wrong input, and the quality drops. A library entry fixes that by storing what made the prompt work and how it fails.
The second reason is change. Vendors update models and features regularly, and a prompt that produced clean output in March may behave differently in October. A stored test input lets you check in two minutes instead of discovering the problem in a client email.
What a library entry contains
| Field | What to write | Example |
|---|---|---|
| Name | Verb, object, version | Draft listing description v3 |
| Owner | One named person | Sales administrator |
| Purpose | When to use it, in one line | First draft from a completed listing sheet |
| Tested in | Tool, plan and date | Team AI plan, October 2026 |
| Variables | The parts that change, in double braces | {{listing_sheet}}, {{word_limit}} |
| Prompt | The full template | See worked example |
| Test input | A fictional input with a known good output | 3-bedroom weatherboard, fictional address |
| Known limits | How it fails and what to watch | Overuses “stunning”; invents walking times |
| Data rules | What may and may not go in | No vendor names, phone numbers or sale reasons |
| Review date | When the owner re-tests it | First Monday of each quarter |
Exercise: build your first three entries
- Find three proven prompts. Search your chat history for prompts you have used successfully at least twice on real work. Ignore one-offs.
- Turn the changing parts into variables. Replace the specific details with placeholders in double braces, such as {{client_type}} or {{notes}}. Keep the rules and format instructions fixed.
- Fill in every field. Use the table above. Known limits matter most: write down what you always fix by hand.
- Create a fictional test input. Write a realistic example with no real names or personal details, run the template, and save an output you would accept as the reference.
- Hand it to a colleague cold. Ask them to use the entry without your help. Note every question they ask and add the answers to the entry.
- Publish and schedule. Put all three in one shared location, link them from wherever your team starts work, and set the review dates in a calendar.
Worked example: a Tauranga real estate office
Illustrative example: a 12-person real estate office in Tauranga has three salespeople writing listing descriptions from the same listing sheet, with very different results. The Real Estate Authority’s guidance is blunt about the stakes: licensees remain responsible for their services, and an error made by a generative AI tool “will not be a defence” (REA). REA’s November 2024 guidance also points to the Fair Trading Act prohibition on false and misleading statements (REA PDF). So the office’s library entry is built around not inventing features.
NAME: Draft listing description v3 OWNER: Sales administrator
TESTED IN: team AI plan, 14 October 2026 REVIEW: first Monday of January
PROMPT
Write a property listing description for a New Zealand residential sale.
Use ONLY the facts in the listing sheet below. Do not describe schools, zoning,
views, renovations, walking times or investment potential unless the sheet states them.
If a fact a buyer would expect is missing (such as heating or parking), add a line
at the end headed "CHECK:" listing it. Do not fill the gap.
Tone: clear and warm. Avoid: stunning, nestled, boasting, must-see, dream home.
Length: {{word_limit}} words. Start with the strongest stated feature.
LISTING SHEET:
{{listing_sheet}}
KNOWN LIMITS: tends to call any kitchen "modern"; check every adjective against the sheet.
DATA RULES: no vendor names, contact details or reasons for sale in {{listing_sheet}}.
Example output from the stored fictional test input (illustrative and abridged):
Sun-filled three-bedroom weatherboard home on a flat 640 square metre section,
with a north-facing deck off the open-plan living area. The kitchen was updated
in 2019 and has a walk-in pantry. Single internal-access garage.
CHECK: heating type not stated; bathroom count not stated.
Before v3, the template had no “CHECK:” line and no banned-word list, and one salesperson’s draft described a heat pump the house did not have. That failure is now recorded under known limits, which is the point of the field.
Where prompt libraries go wrong
- No owner. Entries without a named owner are never reviewed. Prompts quietly break after a model update, and people stop trusting the library.
- Personal data in templates. Copying a real client email into the test input turns the library into a store of personal information. Use fictional inputs only.
- Too many entries. Fifty prompts nobody can find are worse than 12 that everyone knows. Retire unused entries.
- Copying between tools untested. A prompt tuned in one assistant can behave differently in another. Record the tool, and re-test before switching.
- Confusing a library with quality control. A good template lowers the error rate; it does not remove review. Every output still gets a human read.
Engineering teams take this further with version control and automated evaluations; our guide to versioning prompts and evaluations explains how. For a small team, a shared document in SharePoint or Notion works. The AI prompt library template gives you the fields ready to copy, and the person who maintains it is often your internal AI champion. Prompts this good still need careful inputs; the structure behind them is in writing useful business prompts.
Check your library
- I have three entries, each from a prompt that worked at least twice on real work.
- Every entry has a named owner, the tool and date it was tested, and a review date.
- Each entry has a fictional test input and a reference output I would accept.
- Known limits describe what I always fix by hand.
- Data rules say what must never go into the variables.
- A colleague used one entry without my help and I added their questions to it.
Next lesson: Protecting customer information when using AI covers the data rules every library entry depends on. Teams who want a library built during a working session can do it in our AI training for business teams.
Sources
- Generative AI guidance, Real Estate Authority, accessed 19 October 2026
- Generative AI guidance for real estate licensees (PDF), Real Estate Authority, November 2024
Written by the Levy Street team in Wellington. Last updated 19 October 2026.