Skip to content
QDNALearn AI, from beginner to expert
FR

Lesson 18 · Advanced · 20 min

Building your team's prompt library

Build your team's prompt library: one card per prompt with variables, a sample output, known limits and an owner, stored in a shared Page.

By QDNA · Published on · Updated on

Goal
You will be able to turn a proven prompt into a reusable card, with variables in brackets, a sample output and limits, and organise your team's library in a shared Page.
Skills
Frame
A shared Page showing three prompt cards side by side, each with its name, its version, its variables in brackets and a sample output.
Illustration generated by AI

Your first attempt, unaided

Send a colleague, with no word of explanation, the prompt you use most often in Copilot Chat. Ask them to use it, then write down everything they did not understand.

In brief.

A prompt that works for you only works for the team if it is documented: what it does, what to replace, what the result looks like, where it fails. The team library is a set of short cards, one per prompt, with variables in brackets and a sample output, kept in a shared Page that everyone can read and improve. Start with five tested cards, not thirty.

  1. 1What a prompt card must contain

    Microsoft sets the example with its prompt gallery, organised by role and by application. Its help page on writing a good prompt gives the structure: goal, context, source, expectations. A team library applies the same logic to your own tasks, with your vocabulary and your formats. It serves three purposes: reproducibility between colleagues, quality that is measured instead of felt, and handover to newcomers without re-explaining every prompt.

    A card fits on one screen and has six headings:

    Heading Content
    Name and version "Meeting-report v2", date of the last change
    Use case When to use it, when not to
    Full prompt The four ingredients; variable elements in brackets: [audience], [attached file]
    Sample output A fictional, complete result obtained with the prompt
    Known limits What fails: multilingual meetings, files without headers
    Owner Who answers questions and updates the card

    The card's prompt follows the rules of the course: instruction first, source last, a stated length, a rule for missing information, a proofreading instruction. The guides of the model publishers converge on the sample output. Anthropic recommends providing examples to stabilise a format. That example also serves as a reference for the user, who compares what they get with what they should get.

    Where to store the library? In a shared Page that the team co-edits, included in Copilot Chat with SharePoint or OneDrive storage. Or in a document on the team site. Not in Copilot's memory. The persistent memory documented by Microsoft targets the paid licence. According to the documentation in force in September 2026, it is not guaranteed for Copilot Chat without a licence, and it stays personal in any case. Finally, no card contains real data: the examples are fictional, the names and amounts are variables.

    Diagram "The team's prompt library": A colleague; Shared library; Card: context, prompt, expected result; Known trap and fix; Author, date, version; Tested on a real case; Time saved for the whole team. Note: An untested card does not enter the libraryDiagram "The team's prompt library": A colleague; Shared library; Card: context, prompt, expected result; Known trap and fix; Author, date, version; Tested on a real case; Time saved for the whole team. Note: An untested card does not enter the library
    Diagram "The team's prompt library"Diagram generated by AI and reviewed

  2. 2A meeting-minutes prompt that drifted in six months

    A team of five has been sharing by email, for six months, the prompt that writes their meeting reports. Everyone has changed it their own way, and the reports no longer look alike.

    Weak prompt, as it circulates:

    Write the meeting report.

    Each colleague gets a different structure. Owners are invented when the notes give none, and lengths range from ten lines to two pages.

    Strong prompt, as a card:

    Name: Meeting-report v2 (10/09/2026). Owner: the department assistant.
Use case: report of an internal meeting from notes taken during the session. Do not use for a meeting with a client.

Prompt:
Write the report of the [meeting name] meeting of [date] for [audience].
Format: three sections, "Decisions taken", "Actions" (one bullet = one action, one owner, one deadline), "Points in disagreement or open". 250 words at most, neutral tone, in English.
If an owner or a deadline is missing from my notes, write "to be specified" instead of filling in. If no decision was taken on a topic, say so.
Before delivering, proofread: every decision and every action appears in my notes.
Notes: [pasted text]

Sample output: [fictional 200-word report obtained with this prompt]
Known limits: notes in several languages poorly handled; notes without indication of who is speaking.

    The five colleagues get the same structure. The "to be specified" boxes replace invented owners, and a newcomer uses the card from their first day.

    What changes: the variables in brackets say exactly what to replace, so no more wild rewriting. The sample output fixes the format better than any description. The known limits prevent using the prompt where it fails, and the owner receives the feedback.

  3. 3Turn a weekly prompt into a documented card

    Choose a prompt that you or a colleague rewrite every week: progress update, client follow-up, standard reply, monitoring. Here is the starting prompt, in the state it often circulates:

    "Do the project progress update from my notes."

    Rewrite it as a card, with its six headings. Name and version, use case, full prompt with the four ingredients and the variables in brackets. Then a fictional sample output, the known limits and the owner. The composer below produces the card's prompt; enable the "Use the vocabulary of the trade" and "Invent nothing" boosters. Then have a colleague use the card, with no spoken explanation, and note what blocked them. If no colleague is available, reopen the card 24 hours later and run it yourself, adding nothing from memory. Finally, place the card in a Page shared with the team.

    Self-assessment grid: (a) the six headings are filled in and no real data appears; (b) the card was run without help, by a colleague or by you 24 hours later; (c) at least one limit comes from a real attempt.

    Open the prompt composer

  4. 4The card that travels without its use case

    The "Overdue reminder v3" card has worked for six months in the accounts department. It writes a firm reminder, cites the statutory payment term between businesses and announces late penalties. A colleague from the team handling consumer subscriptions copies it from the shared Page. He fills in the variables and sends twenty reminders. The letters threaten penalties that do not apply to consumers, and two customers dispute them. What should have been seen: the card had been tested on a single type of customer, and its "Use case" heading did not say so. Taken out of its context, the prompt produced a wrong result elsewhere. Correction: the "Use case" heading states when not to use the card, for instance "business customers only, never for consumers". The context appears as an explicit variable, [customer type] and [applicable terms and conditions], rather than as an unspoken assumption. The "Known limits" heading records the incident. The card's owner creates a "consumers" variant, tested on five cases before publication. Rule to remember: a card travels without its author, so the context must travel inside the card.

  5. 5Quiz

    Three questions, instant feedback. Each option comes with an explanation.

    1. What makes a prompt reusable by a colleague who has never spoken to you?

    2. An HR team wants to share its job advert prompt, which contains a candidate's name and the salary offered last time. What should be done before putting it in the library?

    3. Review: does giving Copilot a role ("you are a management controller") make its figures more reliable?

  6. 6Proof of mastery

    Paste your complete card (name and version, use case, prompt with variables, fictional sample output, limits, owner), then the written feedback of the colleague who used it without your help and the answer they obtained. If you are learning alone, replace that feedback with your own run of the card 24 hours later, adding nothing from memory, and the list of the points you found missing.

    Advanced badgeThis lesson counts towards the Advanced badgeSee the four badges

    Criteria

Going further

The executive assistant and human resources cards of the site's library are built on this model. The lesson Writing a good prompt: the four ingredients recalls the structure of every prompt, and Pages and Notebooks: keeping and sharing your work explains how to host the library.

Frequently asked questions

Can Copilot memorise our team prompts for us?

Do not count on it. Microsoft documents a memory feature for Microsoft 365 Copilot, but its inclusion in Copilot Chat without a licence is not guaranteed according to the documentation in force in September 2026, and a personal memory cannot be shared. The library lives in a document the team co-edits: a Page, or a file on your team site.

How many prompts do you need to start?

Three to five, chosen among the weekly tasks that several people do the same way: meeting report, follow-up, progress update, standard reply. A library of thirty cards nobody has reviewed is worth less than a library of five cards tested by a colleague.

Do you need a sample output in every card?

Yes. The guides of the model publishers converge: an example of the expected result stabilises the format better than a list of adjectives. The example also serves as a check for the user, who compares what they get with what they should get.

Sources