
Good prompt engineering training teaches your team to hand an AI model the context, the task and the output format it needs, then to test and refine the result until it is reliable. The training that pays off goes one step further: it practises on your own tasks, sets rules for what data may enter which tool, and ends with the prompts that work being turned into repeatable workflows. A one-off lecture on “tips and tricks” does none of that.
This guide is for founders, owners and team leads who want their people to get real work out of tools like ChatGPT and Claude. It covers what a useful curriculum contains, how to pick a format, what to protect, and how to tell whether the training worked.
TL;DR
- Prompt engineering is a small set of learnable habits: give context, state the goal, fix the output format, iterate, and save what works.
- Training sticks when participants work on their own emails, offers, tickets and documents, not on generic examples.
- Decide before the session which data may go into which AI tool. In Switzerland the data protection regulator has publicly addressed generative AI, so this is part of the training, not an afterthought.
- Measure the result in minutes saved and rework avoided on two or three named tasks, then automate the prompts your team uses daily.
- We run AI masterclasses in person or remotely, and we build the automation that follows from them.
What prompt engineering actually is
A prompt is the instruction you give an AI model. Prompt engineering is the practice of writing those instructions so the answer is useful the first or second time, not the seventh. It is less a technical discipline than a communication one: you are briefing a very fast assistant that knows a lot in general and nothing about your company unless you say so.
The model makers describe it in the same plain terms. OpenAI’s help center tells users to be clear and specific, with enough context for the model to understand the request, and to treat prompting as iterative: start with a first prompt, review the answer, then adjust wording, add context or simplify. Anthropic’s documentation for Claude covers the same ground with more advanced techniques such as examples, XML structuring and step-by-step thinking.
That means the core is teachable in a day. What takes longer is making it a habit, and making it fit your business.
What a good curriculum covers
If you compare agendas, the strong ones share the same backbone. Use it as a checklist when you evaluate any training, ours or anyone else’s.
1. The anatomy of a prompt
Participants learn to split a request into parts:
- Context: who the audience is, what the situation is, what the model should know.
- Task: one clear action, such as “summarize”, “classify”, “rewrite” or “compare”.
- Format: a table, three bullet points, JSON, a 120-word email.
- Constraints: what to leave out, which tone to use, which sources to rely on.
A worked example. Weak prompt: “Write a reply to this customer.” Stronger prompt: “You are a support agent at a Zurich software company. Reply to the customer email below in a friendly, direct tone, in under 120 words. Confirm that we received the bug report, give the next step, and do not promise a date. Email: [pasted text].” The second version takes thirty seconds longer to write and usually saves several rounds of correction.
2. Iteration instead of one-shot luck
The most common beginner error is to judge a model by its first answer. Trained users treat the first answer as a draft: they point out what is wrong, add the missing context and ask again. The OpenAI guidance quoted above makes the same point, and it is the habit that separates occasional users from people who rely on the tool daily.
3. Examples and structure
Showing the model two or three examples of the output you want (often called few-shot prompting) is one of the most reliable ways to get consistent results. Asking it to reason step by step helps on analysis tasks. Delimiters such as tags or triple quotes keep your instructions apart from the text the model should process. Anthropic’s guide covers all three for Claude models, and the same ideas carry over to other tools.
4. Testing a prompt like a small piece of software
A prompt that works on one input may fail on the next. A good workshop includes an exercise where you run the same prompt on several different inputs, note where it breaks and fix the instruction. Participants should leave with a simple review standard: does it give consistent results across inputs, does it handle the common edge case, can a colleague read it in half a minute, and does it say honestly what it cannot do?
5. A shared prompt library
A prompt that lives in one person’s chat history helps one person. Saved, named and versioned prompts help the whole team: the offer template, the meeting-notes summary, the tender screening checklist. Training should end with the beginnings of such a library, with an owner who keeps it current.
6. Data handling and limits
Every session should cover what must never be pasted into a public tool, how to spot a confident but wrong answer (a hallucination), and who checks output before it reaches a customer. More on this below.
Practise on your own work, not on demos
The biggest difference between training that changes behavior and training that is forgotten by Friday is the material. Generic exercises (“write a poem about coffee”) teach the interface. Exercises on your real inputs teach the skill: a long customer complaint, a supplier quote, a draft contract summary, a product description in three languages, which matters in a country with four national languages and many bilingual teams.
Here is a way to prepare, whoever delivers the session:
- Collect five to ten real tasks your team repeats weekly. Remove personal and confidential details beforehand.
- Rank them by time spent and by tolerance for error. Internal summaries tolerate more error than customer-facing replies.
- Choose two or three as the training cases. Each participant brings one.
- Agree on what “good” looks like for each case before anyone writes a prompt: length, tone, format, required facts.
One technique worth showing to advanced users is working with long source material. For example, turning a recorded webinar or a competitor’s product video into structured notes is a typical prompt-plus-input task, and tools that offer AI summaries of video transcripts let a team practise extraction prompts on material they already watch.
Who in the company should attend
Not everyone needs the same depth. A useful split:
| Group | What they need | Typical outcome |
|---|---|---|
| Everyone who writes, summarizes or researches | Prompt anatomy, iteration, data rules | Faster first drafts, fewer rewrites |
| Power users (operations, marketing, finance, support leads) | Examples, structured output, reasoning prompts, testing | Reliable prompts for recurring tasks |
| Owners and team leads | Governance, data policy, where AI fits the process | Clear rules, a list of tasks to automate |
| Technical staff or the people building products | Prompts inside software, evaluation, integrations | AI features that behave predictably |
Small teams often combine the first three in one session and handle the technical layer separately. If you are a founder, you usually sit in the third row and the fourth one belongs to whoever builds your product.
Choosing a format
Training comes in several shapes. The right one depends on team size, how much the team already uses AI and how much of the content is specific to you.
- Short introduction (one to two hours). Good for creating awareness and a shared vocabulary. It will not change daily habits on its own.
- Half-day workshop. The practical minimum for hands-on practice: anatomy of a prompt, two or three exercises on real tasks, a first library.
- Full day. Room for advanced techniques, testing and a data-handling session. Suits teams where AI is already in daily use.
- Follow-up over several weeks. Habits form with repetition. A short check-in after two to four weeks, where people bring prompts that failed, locks in more than a longer first day.
On delivery, in person works best for exercises where colleagues compare prompts side by side, while remote works well for distributed teams and short follow-ups. Both are workable if the group is small enough for everyone to get feedback. Our AI masterclasses run in person or remotely, so the format can follow your team rather than the other way round.
For price, ask any provider for a written quote that states the group size, duration, whether the content is tailored to your tasks and whether follow-up is included. Public price lists on the Swiss market vary a great deal because they bundle different things; a lower price for a generic half-day and a higher one for a tailored session are rarely comparable.
Protect your data before you practise
Prompting means typing information into a tool. For a Swiss business that raises a question your training should answer in writing: which data may go into which tool?
The Federal Data Protection and Information Commissioner (FDPIC, in German EDÖB) is the national supervisory authority for data processing by federal bodies and private controllers. On the European Data Protection Day, 28 January 2026, it published a short statement on the use of generative AI and the challenges for data protection. Treat it as a reminder that customer, employee and health data are not casual input for a chatbot.
A practical data-release rule has three tiers:
- Green: public information and your own marketing text. Any approved tool.
- Amber: internal documents without personal data. Only company-approved tools and accounts.
- Red: personal data, contracts under NDA, source code with credentials, financial statements not yet published. Not entered into external tools without a decision from the person responsible for data protection.
Write this on one page and use it in the training exercises. Participants should practise anonymizing a real document (replace names with roles, amounts with ranges) before they prompt on it. If your business handles personal data at scale, check your setup with your data protection officer or a lawyer; the training should support that decision, not replace it.
Check the output, every time
Models can produce fluent text that is simply wrong. Training should install three reflexes:
- Facts and figures are verified against the source, especially numbers, names, legal references and quotations.
- Customer-facing text has a named human reviewer. The prompt creates the draft, a person owns the send button.
- Prompts ask for uncertainty. Instructing the model to say when information is missing or to mark assumptions reduces confident guessing.
These are small rules, and they are what make AI use acceptable to a cautious client, an auditor or a board.
How to tell whether the training worked
Avoid measuring satisfaction scores only. Pick measures that connect to work:
- Time per task before and after. Time five tasks before the session and again four weeks later. A drop of a few minutes per task across a team of ten adds up quickly.
- Rework rate. How often does a first AI draft need a full rewrite? It should fall as prompts improve.
- Library use. How many saved prompts are used each week, and by whom?
- Incidents. Zero cases of restricted data entering the wrong tool.
Set expectations honestly: some trainers report that teams improve immediately after one training day and that mastery builds over several weeks of daily practice with a prompt library, which matches the logic of habit-forming. Plan a follow-up rather than expecting one day to do everything.
From good prompts to automated workflows
Here is the point most training pages skip. Once a prompt works reliably, copying and pasting it by hand is the next bottleneck. The prompt that summarizes every incoming supplier quote, or classifies support emails, or drafts the weekly report, can run by itself inside the tools you already use: your CRM, your shop, your ERP, your inbox.
A simple rule of thumb for deciding what to automate:
| Signal | What it means |
|---|---|
| The same prompt is used by several people every week | Candidate for a shared, automated workflow |
| Inputs and outputs always have the same structure | Easy to connect to a system through an API |
| A mistake is cheap to catch with a human review step | Safe to automate with a review gate |
| The task needs judgment on each case | Keep it as an assisted manual task |
This is where a team’s prompt skills meet software engineering: connecting systems, handling failures, logging what the model did, keeping the whole thing maintainable when the model or your process changes. If you want a prompt library to become a working process, that step is worth planning from the start of the training, not after it.
Where we fit
Ampersand is a small Zurich studio that builds websites, web applications and platforms, and offers AI masterclasses, AI automation for business processes, and integrations with shops, CRM and ERP systems. Senior engineers are involved from the first conversation to handover, so the person who teaches your team what a good prompt looks like is close to the people who can turn the best ones into a working, connected workflow. That is a practical advantage: the training and the automation come from one team, with no hand-offs between them.
You can see the range of what we offer, from automation to integrations and ongoing support, on our pricing and services page.
Questions to ask any provider before you book
- Will you use our real tasks and documents, or a standard deck?
- How do you handle data protection in the exercises themselves?
- Which tools will participants use, and do they need accounts beforehand?
- What do participants leave with: a prompt library, a data-release one-pager, a follow-up session?
- What happens after the session if we want to automate the best prompts?
The answers show quickly whether you are buying a lecture or a change in how your team works.
FAQ
How long does prompt engineering training take?
The basics fit into a half-day workshop, and a full day gives room for advanced techniques and testing. Habits form over the following weeks, so a short follow-up session after two to four weeks makes the biggest difference to long-term results.
Do my employees need technical skills?
No. For most roles prompt engineering is clear writing plus iteration. Technical knowledge only matters when prompts are built into software, such as automated workflows connected to a CRM or shop system.
Is prompt engineering training worth it for a small team?
Yes, when it focuses on the tasks your team repeats every week. Even a small team saves noticeable time when drafts, summaries and classifications come out right on the first or second try, and a shared prompt library keeps that gain when people change roles.
What should we never enter into an AI tool?
Decide this in writing before the training. As a baseline, keep personal data, confidential contracts and credentials out of external tools unless your data protection officer has approved the setup, and anonymize documents before using them as practice material.
Can we turn our best prompts into an automated process?
Yes. Prompts with a stable input and output structure can be connected to your existing systems through integrations, with a human review step where mistakes would be costly. We build this kind of AI automation alongside the masterclasses.
Sources
Talk to us
Have a project this touches on?
A free 10-minute call is the fastest way to find out whether we are the right studio for it.
Book a free 10-min callKeep reading
More articles.
Application Support Models: How to Choose the Right One
In-house, co-sourced, outsourced or hybrid: how application support models work, what each tier handles, and how to pick the one that fits your team.
Read articleIntegration Architecture Patterns: How to Choose the Right One
Point-to-point, API-led, queues, pub/sub and event-driven integration compared in plain language, with a simple way to pick the right pattern for your systems.
Read articleLLM Data Privacy for Swiss Businesses: What to Check Before You Deploy
How LLMs expose business data, what Swiss data protection law expects, and the contract, technical and monitoring checks to run before and after you deploy.
Read article