Everybody in the team re-invents the same prompt for the weekly report, the customer reply or the meeting summary, and the good ones disappear in chat histories. A prompt library turns the best prompts into shared templates with placeholders, an example of good output and a named owner. Typical benefit: more consistent results, faster onboarding of colleagues and less pasting of confidential data because the prompt says what to leave out.
Problem & value
- Today: prompts live in private chat histories, Notes apps or on sticky notes. Results differ by person, newcomers start from zero, and nobody knows which prompt is safe for which data.
- AI value: a library makes recurring tasks repeatable (same structure, same tone, same checks) and spreads the know-how of the two or three colleagues who are good at prompting. It is also the cheapest AI literacy measure: people learn by reusing good examples.
- Benefit figures: this guide does not quote time savings, because no reliable independent figure exists for prompt libraries as such. Measure it yourself: minutes per task before and after, and how many colleagues reuse a prompt.
Variants
- Personal: a note, text expander snippets or saved prompts in the chat tool.
- Team: a shared wiki page, SharePoint list, Nextcloud folder or the sharing function of the AI tool.
- Company-wide: curated by an AI champion network with review and an approved/not approved status.
- Private use: saved instructions and custom commands in the chat app or in a local tool such as LM Studio or Open WebUI.
How it works
- recurring task spotted
- prompt written and tested on anonymised sample
- template with {{placeholders}} + example output + "do not paste" note
- saved in library (owner, version, tags)
- colleague picks template, fills placeholders, checks result
- feedback / improvement
- new version; unused prompts retired
- Key components: a general LLM (any chat tool), a storage place with search and sharing, and a simple review rule. No model training is involved.
- Human in the loop: the user always checks the output; the owner approves changes to a shared prompt.
- Template parts: role and goal, input description, steps, output format, tone, and a sentence on what must not be included (e.g. no personal data of customers).
Data preparation
Which data: the prompts themselves, 3-5 anonymised sample inputs per prompt, one or two examples of good output, and house rules such as tone of voice, terminology and forms of address.
Where it lives: private chat histories, e-mails and Teams chats, Word or Excel files, and the heads of the AI-savvy colleagues.
Typical gaps: prompts that only work with a customer name or real data pasted in, no examples, no owner, several copies with different versions, and sample texts that contain personal data.
Preparation steps: 1. Run a one-hour collection round: each person brings their 2 best prompts (GUIDE-001). 2. Replace real names, amounts and IBANs in samples with dummy values (PREP-022). 3. Add the company vocabulary and tone rules once, centrally (PREP-015, PREP-016). 4. Agree on tags (task, department, data sensitivity) and an owner per prompt.
Tools: a Word/OneNote page, a SharePoint list or Nextcloud folder, Excel for the first inventory. Effort: about 1-2 days for a first library of 15-20 prompts, spread over the team. See GUIDE-003.
Implementation paths
Which path for whom
- SMB - Path A: Microsoft 365 or Google Workspace already offer saving and sharing of prompts, so no new system is needed; a wiki page is enough otherwise.
- Enterprise - Path C: a shared internal AI platform (e.g. Open WebUI) gives group-based sharing, versioning and central governance for many teams.
- Developer - Path C: prompts kept as versioned files in Git next to the code, served by open-source tooling.
- Private - Path A: saved instructions in the chat app of your choice; a local app if data should not leave the device.
Path A - Buy (off-the-shelf)
- Microsoft 365 Copilot: the Copilot Prompt Gallery lets licensed users save prompts and share them with a team or by link; Microsoft's support page states saved prompts can be used, shared and removed but not edited, so a changed prompt is saved again as new version [1].
- Google Workspace: Gemini saved instructions and shared custom assistants (see AUTO-022); check the current admin settings of your edition.
- No licence needed: a Notion/Confluence/OneNote page, Nextcloud (EU-hosted, self-operated or by a German provider), or a text expander tool.
- EU-hosted options: Nextcloud, Confluence/Notion with EU data residency (verify in contract), Mistral Vibe (formerly Le Chat, French provider) with agents; confirm the plan tier.
- Choose A when the team already has an AI tool and needs sharing, not a platform.
Path B - Build on cloud APIs
- A small internal form or n8n/Power Automate flow that fills a template, calls an EU-region model API and returns the result. Worth it only when the prompt is part of a fixed process with own data; otherwise it is over-engineering for this use case.
Path C - Self-host (open source)
- Open WebUI: prompts as slash commands, sharing per user or group, custom input fields, tags and version history with rollback [2]. Runs on a small server (HW1); models can be local (Ollama, small LLM on HW2) or EU APIs.
- Alternatives: LibreChat, Langfuse prompt management for developer teams, plain Git repository.
- Ops effort: about 1-2 days to set up with SSO, then a few hours per month for updates and access rights.
Path D - Custom training
Not relevant. A library replaces training by good examples; fine-tuning only becomes interesting if hundreds of tested prompts and outputs exist for one stable task (see CODE-018).
State of the art (2026)
- Chat tools now ship built-in prompt galleries, saved instructions, projects and shared custom assistants; the standalone prompt-library app category is shrinking [1].
- Best practice: short templates with explicit output format and a good and a bad example work better than long "magic" prompts; stable repeated tasks are better moved into a custom assistant (AUTO-022) or a flow.
- Limits: prompts age when models change; the same prompt can give different results on another model; a shared prompt can carry hidden risk (e.g. pasting confidential data or following instructions hidden in input text, see SEC-010). Reviews are needed, no library is "set and forget".
Shared resources & combinations
- Same building block (llm-general, office-suite-ai): AUTO-001 formulas, AUTO-002 mail replies, AUTO-017 macros, AUTO-018 scripts and AUTO-027 tutoring can all deliver prompts into the library.
- Next step: AUTO-022 turns a proven prompt into a custom assistant; KNOW-011 and KNOW-018 handle wider knowledge capture.
- Baukasten effect: the same Open WebUI or M365 setup serves many AUTO use cases.
Effort, cost & KPIs
- Path A: 1-2 days for collection and setup, mostly staff time; cost is the existing licence.
- Path C: 1-2 days plus a small VM; open-source software without licence fee.
- Cost drivers: curation time, not technology.
- KPIs: number of prompts with owner and review date, reuse count, share of staff using the library monthly, minutes saved per task (self-measured), number of incidents with confidential data pasted.
People & change
Involve: management (AI policy, budget), department heads (which tasks), end-users (they own the best prompts), IT (access, sharing settings) and the DPO (rules on personal data in prompts). Name two or three AI champions as curators (GUIDE-008).
Works council: depends. A prompt library alone does not monitor staff, but AI tools that log user activity or prompts (e.g. M365 Copilot) can be objectively suited to monitor behaviour, which triggers co-determination under §87(1) no. 6 BetrVG; the BAG holds objective suitability is enough. §90(1) no. 3 requires timely information on AI-supported work processes, and §80(3) allows the works council to call an AI expert. Public bodies involve the Personalrat. Do not analyse "who uses which prompt" per person (GUIDE-004).
DPO / DPIA: sample data must be anonymised; a DPIA is usually not needed for the library itself, but an AVV (Art. 28 GDPR) with the AI vendor is (GUIDE-005).
AI literacy: Art. 4 AI Act requires providers and deployers to take measures that support staff's AI literacy; a prompt workshop plus the library is a practical measure (GUIDE-006).
Pilot and concerns: start with one team and 10 prompts for four weeks (GUIDE-009). Typical concerns: "my prompts are my edge" (credit the author), fear of being replaced (the library supports, humans decide), quality of output (always review).
Risks, regulation & compliance
- EU AI Act: minimal risk; only Art. 4 AI literacy applies. Art. 50 transparency applies if prompts produce public-facing content such as chatbot texts or synthetic media (GUIDE-014).
- GDPR: no personal data of customers or staff in prompts or samples; clear "do not paste" notes per prompt; EU hosting and AVV with the provider.
- Other risks: confidential data in prompts (GUIDE-007), stale prompts, copyright of pasted texts, over-trust in output.
- Mitigations: owner and review date per prompt, approved status, short acceptable-use rule, human review of every output.
Getting started
- Hold a one-hour collection round and pick the 10 most used prompts.
- Anonymise samples, add placeholders, example output and a "do not paste" line.
- Store them where people already work (Prompt Gallery, wiki or Nextcloud) with owner and date.
- Present the library in a 30-minute team session and agree on the acceptable-use rule (GUIDE-007).
- After four weeks, retire unused prompts, count reuse and decide on a custom assistant (AUTO-022) for the top three.
Sources
- Microsoft Support - Share your best prompts with others (Copilot Prompt Gallery) - https://support.microsoft.com/microsoft-365-copilot/share-your-best-prompts-with-others
- Open WebUI documentation - Prompts - https://docs.openwebui.com/features/workspace/prompts
- Microsoft Learn - Copilot Prompt Gallery - https://learn.microsoft.com/en-us/microsoft-365/copilot/copilot-prompt-gallery
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj