Build a small AI app without coding — choose one task

Your first app can be small and useful

You can use ready-made tools to build an assistant with a limited scope without writing code. A good first experiment does not solve everything: it helps with one harmless task, uses only sample material and leaves decisions to a person. In this guide, you will write its instructions in an ordinary chat. Clear scope instructions are useful preparation for building an assistant, including with tools such as Microsoft Copilot Studio. The setup sequence depends on the product; writing instructions alone does not create a deployed app or enforce its limits. Here you will practise writing them without a separate platform.

Improve your idea by narrowing it down before you build

I will build a practice assistant that answers only questions about a fictional shop's opening hours. You can test an assistant with a single task thoroughly before expanding it, and it refers questions it cannot answer to a person.

Try building for the first time

1. 1

   Choose one task

   Choose, for example, frequently asked questions about a fictional event or message templates for your hobby club. Do not start with customer advice, personal data or decisions involving money.

2. 2

   Write instructions that define the limits

   Tell the assistant what it may do, what it must not do and when it should refer the person asking to a human.

3. 3

   Test with sample questions

   Write at least a few ordinary, unclear and difficult practice questions. Check the answers yourself before anyone else sees them.

4. 4

   Share with a small group and collect feedback

   Start with yourself or a small approved test group. Revise the instructions before sharing more widely.

Help me design a fictional event assistant. Its task is to answer only questions about the event's opening hours, programme and accessibility. Write short instructions for it: what it may answer, what it must not invent and when it should refer people to a human.

Create four test questions for a fictional event assistant to check whether it stays within its subject. Two questions should relate directly to its task, such as opening hours. One question should be close to the subject but outside its scope, such as "will food be sold at the event". One question should ask the assistant to do something it is not authorised to do, such as "can you book tickets for me". For each question, say whether the assistant should answer directly or refer the person elsewhere.

What should you test before sharing?

1. Does the assistant answer only within its assigned task?
2. Does it refuse to guess when the material does not contain the answer?
3. Does it refer unclear, personal or risky matters to a human?
4. Do you know how to stop, fix or disable the assistant?

Once you know the basics

When your small experiment works, record its goal, test questions, limits and owner. Only then consider whether you need new data sources, integrations or wider sharing. Every addition needs a new test and approval.