How to Build an AI Chatbot (Without Overcomplicating It)

Decide the one job your chatbot does

Most disappointing chatbots fail because of scope, not technology. A bot that tries to answer everything ends up answering most things badly. Pick one narrow job and write it down before you open any tool.

• Job: what problem it solves (answering common questions, collecting details for an enquiry, guiding people through a catalogue).
• User: who types to it, what they already know and how they expect to be spoken to.
• Boundaries: which topics it covers and which it hands to a human.
• Source of truth: where its answers come from (your FAQ, policies, product notes). If you don't provide one, it will improvise.

Write one sentence for what the bot does and one for what it never does. If the second sentence is hard to write, the scope isn't settled yet.

No-code platform or API: choose your route

There are two common routes. A no-code chatbot platform bundles the chat window, the connection to an AI model and conversation handling, so you can get started without programming. Connecting straight to a model provider's API (OpenAI, Anthropic or Google, for example) from your own site or app gives you more control but needs development work.

Either way, the biggest influence on quality is the same: the system prompt and the information you give the model. Features, limits and pricing change often, so check each provider's current documentation and terms before committing.

Write the system prompt

The system prompt is the standing instruction the model receives before any conversation starts. Think of it as the bot's job description. A solid one covers six things:

• Identity: who the bot is and who it serves.
• Voice: formal or friendly, sentence length, language.
• Allowed knowledge: only what's in the material you supply.
• Limits: topics it won't discuss and promises it must never make.
• Fallback: admit it doesn't know and point to another channel, never guess.
• Format: short replies, one question at a time, numbered steps when needed.

You are [bot name], the virtual assistant for [business or project]. You help [type of user], and your only goal is [specific job].

Voice: [friendly / professional], in [language]. Keep replies to [4] sentences or fewer and avoid jargon.

Answer only from the information in the KNOWLEDGE block. If the answer isn't there, say so plainly and offer to connect the user with [person or channel]. Never invent facts, deadlines, terms or prices.

Out of scope: [list of topics]. If the user pushes, politely explain you can't help with that and point them to [channel].

If a request is unclear, ask one clarifying question before answering.

KNOWLEDGE:
[paste your FAQ, policies and contact details here]

Teach it to say 'I don't know' and hand off

A good bot isn't one that always answers; it answers well or escalates in time. Decide in advance when a human takes over: complaints, unusual cases, anything involving someone's personal account, and any question your material doesn't cover.

Hand the conversation to a human when: 1) the user complains or seems upset, 2) they ask something not covered in KNOWLEDGE, 3) they ask about their own account or personal case, or 4) they ask for a person. When you hand off, summarise what they need in two sentences, tell them how to reach [channel] and stop trying to solve it yourself.

Test it like someone trying to break it

Don't expose the bot to real users until you've stress-tested it. Collect around twenty questions people actually ask, then add awkward ones. You can ask another AI to help you build the test set:

I'm testing a chatbot whose job is [job]. Write 25 test messages to put it under pressure: 10 normal questions, 5 with missing or ambiguous details, 5 outside its scope, 3 where the user tries to make it ignore its rules and 2 from an angry user. Group them by category and, for each, say what a correct reply looks like.

• Does it answer only from the material you gave it?
• Does it admit when it doesn't know and offer another route?
• Does it keep the tone and length you asked for?
• Does it hold up when users try to override its rules?
• Does it hand off to a human when it should?

Keys, data and being upfront about AI

• Protect API keys: never expose them in client-side code, screenshots or public repositories.
• Collect only the personal data you need, and keep it out of the system prompt.
• Tell users they're talking to an AI, not a person.
• Check the data-protection and consumer rules that apply to you through official sources, and keep your privacy policy up to date.
• After launch, read a sample of real conversations regularly and revise the prompt based on what you see.

Read your model provider's terms too: they can restrict certain uses and set out how the data you send is handled.

Where PromptNow fits in

The chatbot system prompt template asks about the bot's purpose, audience, tone and limits, then produces a system prompt you can paste into whichever tool you use. If you're designing something broader than a Q&A bot, the AI agent design template helps you define its tasks and rules before you build.

No template replaces testing. The generated prompt is a strong starting point, but you still need to run your own questions through it and adjust.