Building Your First Agent with Custom GPTs

You build a working agent as a Custom GPT with instructions, knowledge and actions.

Why this lesson matters

You can read about agents endlessly, but you only really learn it by building one. Custom GPTs from OpenAI are the easiest way to make your first agent: no code, no server, no technical setup. Within 30 minutes you have a working agent.

In this lesson you build, step by step, an FAQ agent for your organisation: an agent that answers questions based on your documents and can even carry out actions through API connections.

Step 1: Create a Custom GPT

1Go to [chat.openai.com](https://chat.openai.com)
2Click your profile picture → My GPTs → Create a GPT
3You land in the GPT Builder. Choose Configure (not Create, because we want to set things up by hand)

Going deeper

You need a ChatGPT Plus or Team subscription to create Custom GPTs. Plus costs $20/month.

Step 2: Write the instructions (system prompt)

This is the most important part. Your instructions determine how your agent behaves. Here's a template:

text

You are the customer service assistant of [Company name]. 

Your role:
- Answer customer questions based on the 
  knowledge base provided
- Use a friendly, professional tone
- Always answer in Dutch

Rules:
- ONLY give answers you can back up with 
  documents from your knowledge base
- If you don't know the answer, say: "I'll pass you 
  on to a colleague who can help you further."
- NEVER make up information
- Always name the source (which document) with your answer

Style:
- Short and to the point (max 3 paragraphs)
- Use bullet points when there are several points
- Start with the direct answer, then give context

Watch out

Spend at least 50% of your build time on the instructions. Poor instructions = unreliable agent. Test your instructions with edge cases: what if someone asks a question that isn't in the documents? What if someone asks in English?

Step 3: Upload the Knowledge Base

Keep reading

Leave your name and email address and you can read the rest

You get the whole lesson right away, and every other lesson stays open after that. No password, no confirmation email.

Under Knowledge in the Configure tab you upload files:

PDFs of manuals
Word documents with FAQs
Excel files with product information
Text files with company rules

File type	Works well for	Watch out
PDF	Manuals, brochures	Make sure the text is selectable (no scans)
Word/DOCX	FAQ lists, procedures	Use headings for structure
Excel/CSV	Product lists, prices	Keep the columns clearly named
TXT	Rules, templates	Simple but effective

Tip

Structure in your documents = better answers. Use clear headings, bullet lists and tables. An FAQ document in the "Question: ... Answer: ..." format works very well.

Step 4: Add Actions (optional)

Actions are API connections that let your agent do things, beyond just giving answers.

Examples of Actions:

Create a support ticket in Zendesk
Schedule an appointment in Calendly
Look up an order in your order system
Send an email through your mail server

You add an Action through the Actions section in Configure:

text

1. Click "Create new action"
2. Paste your OpenAPI schema (this describes the API)
3. Add authentication (API key or OAuth)
4. Test the action in the preview

Step 5: Test and iterate

Now the real work starts. Test your agent thoroughly:

1Happy path: Ask questions whose answer is in your documents
2Edge cases: Ask questions that are just outside the documents
3Abuse: Try to make the agent step out of its role ("Pretend you're a pirate")
4Language test: Ask in English when the agent should answer in Dutch
5Failure scenarios: What happens when an Action fails?

Adjust your instructions based on what you find. This is an iterative process.

Tip

Start simple: only a Knowledge Base, no Actions. Make sure the answers are right. Only then add Actions. Every layer you add makes the agent more complex, and so harder to debug.

Example: FAQ agent for your organisation

Say you're building an FAQ agent for a web shop:

Component	Content
Instructions	Customer service assistant, friendly, Dutch
Knowledge Base	Returns-policy.pdf, FAQ.docx, Product-catalogue.xlsx
Action 1	Look up order status through the shop API
Action 2	Generate a return label through the shipping API

Result: customers ask questions in the Custom GPT, get answers from your documents, and can check their order status or start a return straight away.

?Knowledge check

What is the most important part of building a Custom GPT agent?

?Knowledge check

What is the best strategy when building your first agent?

?Knowledge check

What should your agent do when it doesn't know the answer?