Prompt Library for Trustees: Proven Prompts for Accounting, VAT and Correspondence

What a prompt library is

A prompt library is an organised collection of proven text instructions («prompts») with which you use a language model such as ChatGPT for recurring trustee tasks: drafting correspondence, structuring VAT questions, triaging incoming documents or summarising long files. Instead of rephrasing each time, you work from templates that clearly state role, task, context, format and limits.

Important: a prompt is an input, not a professional judgement. The model produces plausible-sounding text from patterns – it does not verify whether a VAT rate is current or whether a document is correctly booked. Professional responsibility remains entirely with you as the trustee.

This article shows prompt patterns that have proven useful in daily work and – just as importantly – where their limits lie. It replaces neither professional duties nor legal or tax advice.

Why this matters for trustees

Trustees are bound by professional duties of care and confidentiality. Client data – names, payroll figures, bank details, documents – is particularly worthy of protection. Since the fully revised Data Protection Act (DSG, SR 235.1; informally «revDSG», in force since 1 September 2023), stricter requirements apply to processing personal data, including the use of AI tools.

In the free (Free) and Plus versions of ChatGPT, inputs are by default used to improve the model. Users can disable this under Settings → Data Controls; however, the opt-out does not apply retroactively to earlier conversations. In summer 2025 an experimental feature also became known that allowed users to actively mark chats as public and search-engine-indexable – without the consequences being sufficiently clear; OpenAI subsequently removed this feature. The risk no longer exists in this form today, but it illustrates that platform features change and that caution when sharing remains warranted.

For business use, ChatGPT Team and Enterprise automatically include a data processing addendum as part of the services agreement – no separate conclusion is required – and business data is by default not used for training there. A configurable Swiss or EU data residency, by contrast, is not available with ChatGPT Team, only with ChatGPT Enterprise, Edu and the API Platform. Verify the assured terms with the provider before signing a contract.

The value of a good prompt library therefore lies not only in time saved but in discipline: standardised prompts make it easier to work consistently without identifying client data, to review outputs and to limit the risk of hallucination.

How a good prompt is structured

Proven prompts follow a simple structure: role (who the model should be), task (what to do), context (anonymised key facts), format (desired output) and limits (what not to do, e.g. no invented figures). The more precise these five elements, the more reliable and verifiable the result.

Correspondence example: «You are an assistant in a Swiss trustee office. Draft a polite, factual reminder to a delayed business partner. Context: outstanding documents, second reminder, friendly tone. Format: short letter, Swiss spelling, no amounts. Do not invent names – I will insert them myself.»

VAT example: «Explain in general and neutral terms which criteria are relevant for distinguishing supply of goods from services under Swiss VAT law. Mark explicitly where a review of the concrete facts is needed. Do not invent rates or article numbers; if unsure, say so.»

Document triage and summary example: «Summarise the following – already anonymised – text in five bullet points and flag open questions. Do not change any figures and add nothing that is not in the text.» The last clause is decisive: it lowers the risk of the model «creatively» filling gaps.

In practice: from prompt to verified result

1. 01Clarify the task: is it language/structure (well suited) or concrete facts/figures (only with source check)?
2. 02Anonymise data: replace client data with placeholders (Company A, Amount X). No identifying details into the tool.
3. 03Choose the tool: for business use, Team/Enterprise with an (automatically included) data processing addendum; where CH/EU data residency is required, Enterprise or an own/local solution – not the free private version and not Team.
4. 04Structure the prompt: state role, task, context, format and limits; add «invent no figures, say if unsure».
5. 05Review the result: verify every figure, deadline and norm against the primary source (estv.admin.ch, fedlex.admin.ch).
6. 06Take responsibility and finalise: insert actual client data locally, sign off the text professionally, document source and review.
7. 07Save the prompt: add the proven template to the library and share it within the team.

When it makes sense to use

Language models are strong at linguistic, structuring and preparatory tasks: drafts for reminders, cover letters or client emails; rephrasing into clear language; breaking a complex question into sub-questions; summarising long, already anonymised documents; or creating checklists and conversation guides.

They are also useful as a «sparring partner»: you can have a draft reviewed, request alternative wordings or ask for general background on a tax or accounting topic – always followed by verification against the primary source.

The precondition in every case: no identifying client data in unsafe tools. Work with placeholders (Company A, Person B, Amount X) and insert the actual data only locally in your protected environment. ChatGPT Team and Enterprise protect against training use and include a data processing addendum; where a regulatorily required Swiss or EU data residency is needed, ChatGPT Enterprise or an own or local solution with its own knowledge base is appropriate – not the free private version and not ChatGPT Team.

When you should not use it

Never enter real client names, payroll statements, bank details, social-security numbers or complete documents into the free, private ChatGPT version. This can breach professional secrecy and the Data Protection Act (DSG) and is not technically safeguarded by the default settings.

Do not rely on the model as a source for concrete facts: current VAT rates, deadlines, statutory articles, booking rules or court decisions. Language models «hallucinate» – they sometimes produce false but convincingly worded statements. Every figure- or norm-related claim must be checked against the official source (estv.admin.ch, fedlex.admin.ch).

No AI output is a final client result. Drafts are raw material: you review, correct and take responsibility. For sensitive or complex cases – tax optimisation, legal assessment, delicate communication – professional judgement and final control must remain with you. This is not legal advice.

FAQ

Related topics

Sources

1. KMU-Portal des Bundes – Neues Datenschutzgesetz (DSG/revDSG) · 2026-06
2. Bundesamt für Justiz (BJ) – Datenschutz (DSG, SR 235.1, in Kraft seit 1.9.2023) · 2026-06
3. Eidgenössische Steuerverwaltung ESTV – Mehrwertsteuer · 2026-06
4. OpenAI – Enterprise privacy (Business data not used for training, DPA, data residency) · 2026-06
5. EXPERTsuisse – Weiterbildungsangebot · 2026-06