AI agents in practice: build your first document agent

> 📥 The full pack is available as a PDF (Danish). Download it at the top of the page — all prompts, poll results and the complete Mentimeter export.

On September 22 I built an AI agent live in front of the audience in Aud 0 at ITU. This is the material from that evening, rewritten so you can build the same agent yourself — step by step, in both Claude and ChatGPT.

All prompts work in both tools. Create a project, paste the instruction, upload one document, and test with the three questions. All demo knowledge is fictional (Fjordby Kommune). Never use real client or personal data when practicing.

---

What the room answered

Before we built, the room voted. The result says something important about where we are:

71% have already worked with AI agents (69 of 97), and no one in the room is without tools. The room was further ahead than most people think. The next step is moving from chat to agent.

---

Step 1: Agent Canvas — fill in the frame before opening the tool

If you can't fill in the boundaries, you're not ready to build.

---

Step 2: The meta-prompt — let AI write the instruction

Run this in a regular chat, and paste the answer in as the project's instruction:

```

You are an expert at writing system instructions for AI agents. Write a precise and

robust instruction for a document assistant that answers questions about rules,

policies and collective agreements based on uploaded documents. The agent's purpose:

to give employees and managers a short, correct answer referencing the section the

answer builds on, and to state clearly when the answer is not in the documents. The

instruction must: define role and scope; only use the uploaded documents and never

general knowledge or applicable law outside them; quote the relevant section

verbatim; distinguish between what the document says and what a manager or HR must

decide; never approve, reject or recommend a specific application; answer in a fixed

format (Short answer, Source, Quote, Not in the handbook, Who decides); and keep a

neutral, friendly tone. Write in second person, in clear bullets, ready to paste

directly as an instruction.

```

Short fallback instruction, if you want to skip the meta-prompt:

```

Role: Document assistant for the employee handbook.

Scope: Answer only from the uploaded handbook. No general knowledge, no legislation

or agreements outside the document.

Sources: State the section for every answer, and quote the relevant sentence verbatim.

When the answer isn't there: Write "Not in the handbook" and refer to HR or the

union representative, as the handbook itself does.

Boundaries: You never approve, reject or recommend a specific application. You state

what the rule says, and who decides.

Format: 1) Short answer (two to three lines). 2) Source (section). 3) Quote. 4) Not

in the handbook (if relevant). 5) Who decides.

Tone: Neutral, friendly, precise.

```

---

Step 3: Test with three questions

Two are in the handbook, one is not. Notice what the agent does when the answer isn't there.

1. May I work from Italy for two weeks in October? Expected: the rule in § 3.4 reproduced, and the decision placed with manager and HR.

2. May I use an AI assistant to write a reply to a citizen? Expected: § 5.3 quoted verbatim.

3. What are the rules for parental leave? Expected: "Not in the handbook" and a referral to HR or the union representative (§ 1.2).

Extra questions: How many days a week may I work from home, and what does it require? Can you approve my home-working agreement? (It must say no.) What does the collective agreement say about overtime pay? (Outside the document.)

---

Step 4: Scheduled runs — the agent works on its own

Write the request in plain language in the project. In Claude it becomes a scheduled task, in ChatGPT a Scheduled Task. Afterwards, check that the task is listed and enabled.

Rule of the day, every morning:

```

Every morning at 07.00: review the uploaded employee handbook and write "Rule of the

day": one section that employees often ask about, explained in three lines with the

section number and a verbatim quote. Pick a new section each day. Add nothing that

isn't in the handbook.

```

Gap report, a one-off task you can watch land immediately:

```

Create a one-off task that runs in five minutes: review the uploaded employee

handbook and produce a "Gap report": list the three topics employees typically ask

about that the handbook does NOT cover, and for each topic refer to who should be

asked instead.

```

A schedule is a time trigger. Event triggers (a new file, a new email, a new form response) are the next step and typically require Power Automate, Zapier or Make.

---

Step 5: The app prompt — one prompt, the whole app

Turn the agent's logic into a tool colleagues can use without knowing AI. Adjust colors and texts, and insert your own sections. The full app prompt (layout, function, ready-made questions and rules) is in the PDF pack — it builds "Regelopslag", a small single-page web app in one file, with no external calls and no backend, that analyzes the handbook text locally in the browser and always answers with section, quote and "Who decides".

---

The same agent in Claude and ChatGPT

The craft is the same. Only the buttons have different names.

---

All demo knowledge (Fjordby Kommune and the employee handbook) is fictional and invented for teaching purposes. The pack is general practical guidance, not legal advice. Check your own rules for approved AI tools and personal data before building on real documents.