Deutsch

Knowledge

For employees in businesses that work with AI or are just starting. Short, without jargon and meant to be passed on. What is written here is what we apply in our own projects.

What a language model does

A language model predicts, word by word, what is most likely to come next. It does not look anything up. That is why an answer often sounds right even when it is not.

Installing ET-4711 takes about

  1. 30 minutes41 %
  2. one hour27 %
  3. 20 minutes14 %
  4. two hours9 %

All four continuations sound plausible. Which one is right is only in the installation guide. If the model is given the guide, it can read the number there and cite the page.

  • A fluent answer is not evidence.
  • Ask for the source. With a source an answer can be checked, without one it cannot.
Simplified, with made-up values

A good instruction for the AI

Most weak answers come from instructions that are too short. Five building blocks help. Switch them on and watch the instruction grow.

A good instruction for the AI
Your instruction

RoleYou work in customer service for a heating manufacturer. ContextAn installer asks whether spare part ET-4711 fits series B systems. The data sheet is attached. TaskWrite a reply to the customer. FormatFive sentences at most, polite form, and at the end the page of the data sheet you rely on. ExampleThis is how we usually reply: “Hello, according to the data sheet (p. 2) ET-4711 fits series B systems.”

If the answer does not fit, say what is missing instead of starting over: “shorter”, “no jargon”, “cite the source”.

Replace confidential details

Most tasks do not need names, addresses, account or invoice numbers. Replace them with placeholders before the text goes into an AI tool. The result does not get worse.

Replace confidential detailsMade-up data

Please cancel invoice RE-2026-0412<invoice> for the Berger family, 4 Linden Road<customer>, <address> and refund the amount to DE00 1234 5678 9000 0000 00<IBAN>.

Which tool may be used with which data should be set down in writing in your business.

A folder the AI works in

An AI agent works in a folder of plain text files. A fixed layout makes sure the agent knows where it is, what to do and where the result goes. Pick a file.

A folder the AI works in
Workspaceworkspace/

One folder for a team or a department. Everything in it is plain text and can be read with any editor.

MapIDENTITY.md

What this folder is and how it is organised. The agent reads the map first.

SignpostCONTEXT.md

Which task lives in which subfolder. The agent only loads the folder it needs right now.

Rules_config/

Tone, for example German and polite form, and data protection rules that apply to every task.

One tasksort-enquiries/

One subfolder per recurring task. For a new task, copy it and adapt the instructions.

InstructionsCONTEXT.md

What goes in, what should come out, which fields the result has and when the agent should ask instead of guessing.

Inputinput/

Put here what should be worked on, such as an email or a few lines of notes.

Resultresult/

The agent writes its proposal here. It sends nothing out. A person checks and decides.

  • One task per run.
  • The agent proposes, you decide.
  • No passwords and no customer data in plain text.

Check before it goes out

An AI result is a draft. Five questions are usually enough. Tick what you have checked.

of 5 checkedReady to send

Templates to download, such as an AI inventory and a usage policy, are in progress.