Chatbot conversation flow designer
Good chatbots are not one big prompt — they are a small set of well-defined nodes that route the user to the right answer or to a human. This tool helps you map that flow. Enter your bot’s purpose, the intents users will arrive with, and the triggers that should escalate, and it produces a dialogue tree with a ready-to-use AI prompt for every node.
How it works
The designer turns your inputs into the standard structure every task-oriented bot needs:
- A greeting node that states what the bot does and offers your intents as quick replies.
- An intent node per user request that confirms understanding, asks the one clarifying question that matters, and sets expectations.
- A fallback node for when the bot does not understand, which apologises and offers a rephrase or a human.
- An escalation node that hands off cleanly to a person when your triggers fire, summarising what has been collected.
Each node carries its own prompt, anchored to your bot’s purpose, so the responses you generate stay consistent across the whole conversation.
Why node-per-intent matters
A single monolithic prompt handling every topic tends to drift — the model loses the thread on long conversations, contradicts itself between turns, or conflates two different intents when the user’s phrasing is ambiguous. Breaking the flow into explicit nodes has two concrete advantages. First, each node prompt is short and scoped, so the model stays focused on one decision at a time. Second, you can measure node-level success rates: if users hit the fallback node repeatedly on a particular intent, you know to improve that specific node, not the entire bot.
Tips and examples
Write the purpose as a concrete job — “help customers track orders and start returns” beats “customer support” — because it flows into every node. Keep intent names in the user’s words (“track my order”, not “order_status_lookup”), since those double as the quick-reply labels. Pay special attention to the escalation node: the fastest way to lose trust is a bot that traps a frustrated user, so list explicit triggers like “asks for a human” or “refund over £100” and let the flow hand off gracefully. Generate the responses, then test them against real example messages before you ship.
Common design mistakes
- Too many intents in the greeting. Offering more than five or six quick-reply options overwhelms users. Pick the top three to five by volume and route everything else to the fallback.
- Vague escalation triggers. “User is unhappy” is not actionable; “user has used the word ‘refund’ twice” or “conversation exceeds 6 turns without resolution” is.
- Missing context handoff. When the escalation node fires, the human agent needs the conversation summary and any data collected. Include a context-capture step in the escalation prompt.
- Skipping the fallback node. A bot with no defined fallback either goes silent or hallucinates an answer. An explicit “I didn’t catch that — could you rephrase?” response keeps the user in the loop.