Prompting in a multilingual company: techniques, templates & the language question (2026)

A prompt is the input with which you give an AI a task — the instruction or question in natural language from which a language model generates its answer. Whether that answer is usable straight away or stays generic is almost always decided by the prompt, not the model. This guide shows what makes a good prompt, the six core techniques with concrete examples, ready copy-and-paste templates for five departments, the most common mistakes — and what changes once real company data enters the prompt.

One thing up front, because it comes up in every multilingual company: the language you prompt in matters far less than the structure of the instruction. Current models are trained multilingually and handle German, English, French and the other major European languages reliably. You do not need to switch to English for quality reasons.

What does matter is consistency. The practical rule for a team spanning languages: prompt in the language the output is meant to be in, and state that language explicitly. A German customer email written from an English prompt goes through an implicit translation step that costs register and idiom. Making the output language part of every shared template is what keeps results consistent no matter who runs it.

What makes a good prompt

A weak prompt reads: "Write me an email to a customer." The model has to guess — which customer, which matter, which tone, which length. A good prompt answers five questions before the model answers:

• Who should answer? (the role)
• What is the background? (the context)
• What exactly should be done? (the task)
• How should the result look? (the format)
• How does the model recognise a good result? (examples)

Remember: the more a person would have to guess to do your task, the more context the model needs too. A good prompt is a good brief.

The six core techniques

1. Role — give the model a perspective

An assigned role steers depth, tone and viewpoint. "You are an experienced tax adviser" produces different answers than "you are a marketing intern".

You are an experienced sales director in B2B machine building.
Phrase the answer the way you would explain it to a new account executive.

Example output (extract): "Before you send the quote, settle three things: budget range, decision-maker and time horizon. Without those three, any quote is a shot in the dark…" — the role shifts the answer from generic advice to practical sales knowledge.

2. Context — supply the situation

The model does not know your company. Sector, audience, constraints and goal belong in the prompt.

Context: we are a mid-sized manufacturer of laboratory equipment (250 employees).
Audience: technical buyers at pharmaceutical companies.
Goal: a first contact by email after meeting at a trade fair.

Effect: with context, the model gets the tone and the domain vocabulary right; without it, you get platitudes that apply to every sector and fit none.

3. Task — one clear verb, one task

Start with an unambiguous verb: create, summarise, analyse, compare, translate, rewrite. One task per prompt — several tasks produce washed-out results.

Task: summarise the following minutes into five key statements
and list open to-dos with their owner separately.

4. Format — structure the output

Say explicitly what the result should look like: table, bullet points, character length, headings. Delimiters such as triple quotation marks (""") or headings separate context from task and prevent confusion.

Format: answer as a table with the columns "Statement", "Priority" and "Next step".
The text to analyse sits between the triple quotation marks:
"""
[insert text here]
"""

Example output:

5. Examples — show rather than describe

When tone or structure is hard to describe, supply an example. A single pattern (one-shot) or several (few-shot) anchor the desired result more reliably than any explanation.

Write product descriptions in the following style.
Example:
Input: "cordless screwdriver, 18V, 2 gears"
Output: "Powers through every screw — the 18V cordless screwdriver with two gears
for precise work from flat-pack furniture to the building site."
 
Now for: "table saw, 1500W, rip fence"

6. Iteration — sharpen rather than start over

The first attempt is rarely perfect. Instead of writing a completely new prompt, correct deliberately in the same conversation:

Good, but cut it to a maximum of 80 words, make the tone a little more factual
and remove the overstatement in sentence 2.

Iteration is the most important and most frequently forgotten technique. It costs seconds and lifts quality more than any longer opening prompt. After a longer conversation it is also worth asking: "summarise the final version cleanly" — that gives you a usable end result without the whole conversation history.

A complete example: from weak to strong

Weak prompt:

Write a summary of our customer meeting.

The result is inevitably pale — the model knows neither the meeting, nor its purpose, nor the desired format.

Strong prompt (all six techniques combined):

You are an account manager who follows up customer meetings in a structured way. [role]
Context: meeting with Muster GmbH, the topic was introducing our platform
in their accounting department. Participants: head of IT and the business unit. [context]
Task: summarise the following minutes and derive the next steps. [task]
Format: first 3 key statements as bullet points, then a table of to-dos
with columns "Task", "Owner", "Deadline". [format]
Mark with [CHECK] wherever a statement is not clearly evidenced in the minutes. [safeguard]
Answer in English.
Minutes:
"""
[insert minutes]
"""

Example output (extract):

• Muster GmbH wants a pilot in accounting, preferred start Q3.
• The main blocker is the IT security review; the business unit is convinced.
• Budget exists in principle but has not been finally approved. [CHECK]

Note the added line: "Answer in English." In a team that works across languages, that single instruction is what keeps a shared template from drifting.

Ready prompt templates by function

These are meant for copying and adapting. Replace the placeholders in square brackets. Each template combines role, context, task and format — and names the output language, because a shared template gets used by people prompting in different languages.

Sales — follow-up after a first meeting

You are an experienced B2B sales representative.
Context: I had a first meeting with [company] about [product/matter].
Discussed: [bullets]. Left open: [open points].
Task: write a follow-up email that names the next steps clearly
and contains a concrete proposed date.
Format: maximum 120 words, friendly-professional tone, with a subject line.
Output language: English.

HR — a job advertisement from key facts

You are a recruiting specialist.
Context: we are looking for [position] for [department] at [location], [full/part time].
What matters: [responsibilities], [requirements], [what we offer].
Task: create an engaging, non-discriminatory job advertisement.
Format: sections "Your responsibilities", "Your profile", "What we offer" as bullets,
plus a motivating opening sentence. Inclusive language.
Output language: English.

Marketing — a LinkedIn post from a specialist article

You are a content marketing manager with clear, unexcited language.
Context: the audience is [audience]. Core message: [one statement].
Task: turn the following article into a LinkedIn post.
Format: hook in the first line, 3 short paragraphs, one question at the end.
No hashtags beyond three. Output language: English.

Support — a consistent draft reply

You are a support agent.
Context: our tone is friendly and solution-oriented. Product: [product].
Task: draft a reply to the customer enquiry below.
If information is missing, list the follow-up questions separately rather than inventing answers.
Format: email, maximum 150 words. Output language: same language as the enquiry.

Analysis — a document reduced to a decision

You are an analyst preparing decisions for a management board.
Task: summarise the document below on half a page.
Format: core message, three most important points, risks, recommendation.
Mark anything not clearly evidenced in the document with [CHECK].
Output language: English.

That last template shows the multilingual case at its most useful: "same language as the enquiry" lets one shared support assistant serve a customer base that writes in several languages, without anyone maintaining a separate template per language.

The most common mistakes

• Too vague, no context. The model fills gaps with generalities.
• Several tasks in one prompt. Each gets done a bit, none properly.
• No format specified. You get prose where you needed a table.
• Giving up after the first answer. Iteration costs seconds and is where most of the quality is.
• Taking output unchecked. Quotes, contracts and figures need human approval before use.
• Sensitive data in an unapproved tool. That is shadow AI, and it is a data-protection problem rather than a prompting one.
• Leaving the output language implicit in a shared template. Fine while one person uses it; inconsistent the moment a colleague prompting in another language runs the same assistant.

When real company data enters the prompt

As long as you are drafting generic text, prompting is a craft question. The moment personal or confidential data enters the prompt, it becomes a data-protection question. What is then required:

• A platform operated in the EU, with a data processing agreement (DPA).
• The contractual assurance that inputs are not used to train the models.
• Central roles and permissions, so each person reaches only the data they are approved for, with access logged.

Private AI accounts do not satisfy any of this. meinGPT is operated in the EU by SelectCode GmbH, is ISO 27001 certified and has its security examined regularly by independent penetration tests; evidence is available through the Trust Center.

From a good prompt to a shared assistant

The biggest lever is not the individual prompt but what happens to it afterwards. A prompt that works and is never shared is a one-off. Captured as a shared assistant — role, context, knowledge base, format and output language fixed once — it becomes a team resource that pays off at every repetition, and everyone works with the same quality-assured tool.

How to select and introduce such a platform is described in AI for companies; the step from individual prompting to team adoption is covered in using ChatGPT effectively.

Conclusion

Good prompting is not a secret technique but a clear brief: role, context, task, format, examples — then iterate. Which language you prompt in matters far less than most teams assume; what matters is stating the output language and staying consistent, especially in shared templates that colleagues across a multilingual company will run. And once real company data is involved, the decisive question stops being how you phrase the prompt and becomes where it is processed.