Writing AI work instructions: prompts in the building process

A work instruction to an AI, known in the trade as a prompt, is like an instruction to a new colleague who does not know how you work. Written well, it names five things: the role, the documents, the task, the form wanted for the result and a check step. Whatever you leave out, the model fills in with an assumption. The result remains a draft that you check against the documents.

Why the instruction has to be so precise

The Anthropic documentation compares the model with a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result. As a test it recommends showing the instruction to someone with minimal context on the task: if they would be confused, the model will be too. The OpenAI documentation puts it similarly: give the model any additional information it might need to generate a response. Both guides are as of 13 November 2026. They apply to the models they describe and change with new versions. Treat the advice as a starting point and test it on your own tool.

How a language model arrives at its answers, and why it can fill gaps without documents, is covered in the article Language models: what they can and cannot do. Here the subject is the side you can influence: the instruction.

The five building blocks

Building block	What goes in
Role	Who the model should be and what the text is for
Documents	Everything the text is to be based on
Task	What exactly should result, in one clear sentence
Form	Layout, length, tone, columns
Check step	References, open points, a note on gaps

On the role: according to Anthropic, even a single sentence sets behaviour and tone. In a chat interface without a system setting, you put it at the start of the instruction. On the reason: if you explain what the text is for, you give the model more information about your goal, and it returns more targeted answers.

On the documents: according to Anthropic, long documents (the documentation names more than 20,000 tokens) are best placed at the top of the instruction, above the actual question. With several documents, the documentation recommends labelling each one with its content and source. In your practice that means giving every document a title, such as “Notes, site meeting 14” or “Contract, section 4”. The OpenAI documentation sets a different emphasis on order and names the end of the instruction as the better place for additional information. More important than the order is that the documents are in the instruction in full and carry a name. Try both orders on a case you know with your own service.

On the form: the documentation asks you to be specific about the output format and constraints. “Short” is not a specification. “At most half a page, three columns” is.

On the check step: Anthropic recommends explicitly allowing the model to say “I don’t know”, and having it extract word-for-word quotes first when documents are long. The model can likewise support every statement with a passage and strike a statement if it finds none. The same page states that these techniques do not eliminate hallucinations entirely and that critical information still has to be validated. That is why the check step belongs in the instruction itself. You ask for the source passage and for gaps to be reported in the instruction, not only after reading the result.

1. Role Who the AI is and what the text is for
2. Documents Attach notes, contract, list in full
3. Task One clear instruction in one sentence
4. Form and check Layout, length, references, open points
5. Draft The AI prepares the text AI prepares
6. Your check Compare with the original, sign off

Text version of the graphic

• 1. Role: Who the AI is and what the text is for
• 2. Documents: Attach notes, contract, list in full
• 3. Task: One clear instruction in one sentence
• 4. Form and check: Layout, length, references, open points
• 5. Draft: The AI prepares the text (the AI prepares)
• 6. Your check: Compare with the original, sign off

A running example with invented data

The data below is invented. The subject is a draft of meeting minutes made from notes. For processing recordings, see the article Meeting minutes with AI.

The instruction could read:

You are an assistant to the site manager on a new-build project (example project “Kita Beispielweg”, invented data). You prepare drafts of minutes that the site manager checks before they are sent.

Document 1, “Notes, site meeting 14”: [paste notes here]. Document 2, “Minutes, site meeting 13”: [paste text here].

Task: From document 1, produce the draft minutes of site meeting 14.

Form: a table with the columns No., Topic, Decision or status, Responsible, Date. Take names and dates over only if they appear in the notes. Write “open” in the cell if a detail is missing, and do not guess.

Check step: For every row, name the passage of the notes it is based on. At the end, list all items from minutes 13 that no longer appear in the notes.

You do not read the result from top to bottom. First check the rows marked “open”, then names and dates against the notes, then the list of items that have disappeared. The AI has saved the writing of the table, not the checking.

Typical mistakes

• Too little material: “Write the minutes of the last meeting”, with no notes. The model then invents plausible items.
• Several jobs in one: minutes, programme and covering letter in one request dilute every result. Split them up.
• No form given: you get running text where you need a table.
• No permission for gaps: without the sentence “write open if it is missing”, the model closes the gap with an assumption.
• Unchecked third-party text: the BSI points out that instructions can also be hidden in web pages, emails or documents that the model evaluates, and that the model may then change its behaviour. Have third-party text, such as incoming emails, evaluated only as material, label it that way in the instruction and read the result critically. That is no reliable protection.
• Instruction never tested: for larger applications the OpenAI documentation advises building tests that measure how an instruction behaves, so that you see the effect as you improve it. In an office, the author’s assumption is that a new instruction tried on a case whose result you already know is enough.

For outputs that have an effect outside the company, the BSI recommends checking them where the consequences may be critical and comparing them with other sources where needed. Minutes that go to the parties involved belong in that category.

What you can do now

1. For a document you need often, write an instruction with the five building blocks and save it as a template.
2. Show the instruction to a colleague who does not know the task. If they ask questions, add the answers.
3. Test the instruction on a closed case and compare the result with the real document.
4. For project data, use only platforms that can be operated in line with the GDPR, never a consumer chat service. Which data may go into which tool is covered in the article Data zones: which data may go into which tool?. Clarify this beforehand with your data protection officer.