Chatbots & Conversational AI
System prompts, rules & conversation design
The Problem: You want a support bot, but it forgets context after 2 messages, answers questions about cooking instead of your product, and promises refunds it cannot authorize. How do you make it reliable?
The Solution: A Waiter, Not a Know-It-All
A chatbot is an LLM wrapped in a system prompt that defines its role, rules, and boundaries. Think of a waiter: they know the menu, remember your order, and politely decline to give medical advice. The three key ingredients are: role (who is the bot), rules (what it can and cannot do), and conversation history (so it remembers context). Without rules, even the best LLM will go off-topic, hallucinate, and break character.
Think of it like a waiter at a restaurant:
- 1. Define the role: Who is this bot? A support agent, a tutor, a sales assistant? The role sets the tone and expertise area
- 2. Add behavioral rules: Stay on topic, admit uncertainty, limit response length, refuse harmful requests — each rule is a guardrail
- 3. Test on edge cases: Off-topic questions, jailbreak attempts, out-of-scope requests — test what happens WITHOUT each rule
- 4. Iterate and harden: Every failure reveals a missing rule. Add it, test again, repeat until the bot handles all scenarios gracefully
Where Are Chatbots Used?
- Customer Support: Answering FAQs, handling complaints, escalating complex issues
- Personal Assistants: Scheduling, reminders, task management in natural language
- Tutoring & Education: Explaining concepts, guiding through problems without giving answers
- Sales & Onboarding: Product recommendations, guided setup, FAQ automation
- Common Pitfall: No Boundaries: A chatbot without clear rules will answer any question, make up facts, and promise things it cannot deliver — always define what the bot should NOT do
Fun Fact: The first thing users try with any chatbot is breaking it. "Ignore your instructions" attacks (jailbreaking) are so common that production bots need multiple defense layers. A single-line "You are a helpful assistant" system prompt survives about 3 creative user messages before the bot starts roleplaying as a pirate.
Try It Yourself!
Use the interactive constructor below to build a chatbot step by step and see how each rule changes its behavior.
Choose a role for your bot
Frequently asked questions
What is a system prompt and why is it needed?
A system prompt defines the chatbot's persona, capabilities, constraints, and behavior rules. It's the 'operating instructions' that shape every response. Without it, the LLM has no consistent personality or boundaries.
How do I restrict a chatbot's conversation topic?
Define allowed topics in the system prompt and add explicit rejection rules. Use topic classification before generating responses for stronger enforcement.
How do I prevent chatbot jailbreaking?
Layer multiple defenses: strong system prompt with explicit refusal rules, input classification to detect manipulation attempts, output filtering, and regular red-team testing. No single defense is sufficient.
How do I manage dialog history and memory?
Use a sliding window of recent messages (last 10–20 turns), summarize older context, and store key facts separately. For long conversations, implement explicit memory storage with retrieval when relevant.
Try it yourself
Interactive demo of this technique
Create a customer support chatbot system prompt for an online store
You are a support assistant. Answer user questions.
You are Alex, a virtual assistant for TechStore online shop. Your goal is to help customers quickly and in a friendly manner.
Your scope:
- Order status and delivery tracking
- Returns and exchanges (within 14 days)
- Warranty information (1-year on all electronics)
- Delivery options and timelines
- Questions about products in the catalog
Tone and style:
- Friendly but professional
- Concise answers (2–4 sentences)
- No technical jargon
- Always offer a concrete next step
If question is out of scope: Say: "A specialist can help better with that. Would you like me to connect you with an agent?"
Greeting at conversation start: "Hi! I'm Alex from TechStore. How can I help you today? 😊"
A good chatbot system prompt defines a role, scopes the domain, sets the tone, and specifies fallback behavior — this prevents hallucinations and user frustration.
Create a free account to solve challenges
2 AI-verified challenges for this lesson
This lesson is part of a structured LLM course.
My Learning Path