Short answer
An AI prompt library template is a shared register of tested prompts that a team reuses for recurring tasks. Each entry records the task, owner, approved tools, data allowed, required inputs, the prompt text, a known-good output, review rules and the date and model it was last tested on. This template includes five starter prompts for New Zealand businesses.
- Store prompts with a known-good output, so anyone can tell when results drift.
- Record which data classes each prompt may use.
- Retest prompts when the underlying model changes.
- Retire prompts nobody has used in three months.
Use this template when the same AI tasks come up every week and different people are getting different results. Start small: five prompts that are used often beat fifty that are not. Our lesson on building a reusable prompt library walks through the first session.
The prompt library template
Part A: Prompt entry
One entry per prompt. Every field is required before a prompt is marked approved.
| # | Field | What to record |
|---|---|---|
| 1 | ID and name | P-01, “Customer email reply draft” |
| 2 | Task and when to use it | One sentence, including when not to use it |
| 3 | Owner | The person who maintains and retests it |
| 4 | Approved tools | The tool and account type it is tested on |
| 5 | Data allowed | Public, internal, or confidential and personal; never more than the tool is approved for |
| 6 | Inputs required | What the user pastes or attaches, with placeholders shown in square brackets |
| 7 | Prompt text | The full prompt |
| 8 | Known-good output | One example output the owner has approved |
| 9 | Review required | Who checks the output and what they check |
| 10 | Version and date | v1.2, 20 October 2026 |
| 11 | Last tested on | Tool, model name and date |
| 12 | Status | Draft, approved or retired |
Part B: Library index
| ID | Name | Team | Owner | Data allowed | Last tested | Status |
|---|---|---|---|---|---|---|
| P-01 | ||||||
| P-02 |
Part C: Library rules
- Prompts contain instructions and placeholders only, never real client or personal information.
- Each prompt has one owner, who retests it when the tool’s model changes.
- A prompt is approved only after the owner has run it on at least five real tasks.
- Users report bad outputs to the owner with the input, if it is safe to share.
- Prompts unused for three months are retired, not left to go stale.
Five starter prompts
Adapt these to your business, test them on five real tasks, and record the results before approving them.
P-01 Customer email reply draft. Data allowed: confidential, in an approved business tool only. Pair it with the review habits in drafting emails with human review.
You draft replies for [business name], a [type of business] in New Zealand.
Write a reply to the customer email below. Use plain New Zealand English,
under 150 words, friendly and direct. Use only the facts in the email and
in the policy text provided. If the policy does not answer the question,
say a team member will follow up by [timeframe]. Do not promise refunds,
dates or prices that are not in the policy text.
POLICY TEXT: [paste]
CUSTOMER EMAIL: [paste]
P-02 Meeting notes to actions.
From the meeting notes below, list each action as: owner, action, due date.
If an owner or date is not stated, write "not stated". Do not invent either.
Then list decisions made, then open questions. Keep each line under 20 words.
NOTES: [paste]
P-03 Summarise a document with page references.
Summarise the attached document for [audience] in no more than 8 bullet points.
After each bullet, give the page number it comes from in brackets.
If something important is unclear or contradictory, list it under "Check".
Do not add information that is not in the document.
P-04 Answer a question from our documents only.
Answer the question using only the documents provided. Quote the sentence
you relied on and name the document. If the documents do not answer it,
reply "Not covered by our documents" and suggest who to ask.
QUESTION: [paste]
P-05 Plain English rewrite.
Rewrite the text below in plain New Zealand English for [audience].
Keep every fact, number and obligation. Use short sentences and NZ spelling.
List any sentence where simplifying could change the meaning.
TEXT: [paste]
How to build the library
- Collect what people already use. Ask each person for their two most-used prompts.
- Pick five high-frequency tasks. Choose tasks done weekly by several people.
- Write and test. Run each prompt on five real tasks in the approved tool. Save the best output as the known-good example.
- Assign owners and publish the index. Put it where the team already works.
- Retest on model changes. Vendors change models often: OpenAI replaced GPT-5.6 Sol and Luna with GPT-6 in ChatGPT on 7 October 2026 (OpenAI), and Anthropic retires Claude Sonnet 4.5 from its API on 30 November 2026 (Anthropic).
Filled-in example
Illustrative example: a 15-person Nelson property manager’s entry for routine maintenance replies.
| Field | Entry |
|---|---|
| ID and name | P-07, “Tenant maintenance acknowledgement” |
| Task | Acknowledge a routine maintenance request and say what happens next; not for urgent or safety issues |
| Owner | Senior property officer |
| Approved tools | Microsoft Copilot Chat, work account |
| Data allowed | Internal; tenant name and address only |
| Inputs | Tenant email; property address; trade type |
| Known-good output | Saved reply acknowledging a leaking tap, with the 3-working-day contact commitment |
| Review | Officer reads in full before sending; urgent issues go to the property manager instead |
| Version, last tested | v1.1, 20 October 2026; tested on five requests on 20 October |
| Status | Approved |
Common mistakes
- No known-good output. Without it, nobody can tell when quality slips.
- Real data in prompts. The New Zealand Law Society’s guidance says client or personal information should not be used for testing AI or generating templates (NZLS), and the same rule suits any business.
- Prompts that hide the review step. Every entry says who checks the output; see the AI quality review checklist.
- Too many prompts. A long library nobody maintains is worse than a short one that works.
Next step
Teach the library in a short session using writing useful business prompts and giving AI useful context. If prompts power a production system rather than chat, manage them like code with versioning prompts and evaluations.
Sources
- GPT-6 in ChatGPT, OpenAI Deployment Safety Hub, 7 October 2026
- Model deprecations, Anthropic, accessed 19 October 2026
- Generative AI guidance for lawyers, New Zealand Law Society, updated 24 February 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.