An agent answer in your own format

A model writes differently every time: sometimes it greets you, sometimes it signs off, sometimes neither. A brand has a frame that should not depend on the model’s mood: a greeting, a signature, a disclaimer, a link to a human. A template separates the two.

What is inside

1. Customer question — the entry.
2. Model’s answer — the agent, answering as usual.
3. One shape for every answer — a template node. It drops the agent’s answer into your text: {{ nodes.agent.output }} means “the output of the node whose id is agent”.
4. Send the answer — the exit.

The reference points at the node’s id, not its caption: rename the node on the canvas and the template keeps working. Change the id itself and you must update the template, or it renders empty and the customer receives a frame with nothing in it.

What to set up for yourself

• The template text. Everything around {{ nodes.agent.output }} is yours: a heading, a signature, opening hours, a link to a human.
• Formatting. The chat on your site and Telegram both render Markdown: **bold**, lists, links. In a channel without formatting the asterisks arrive as asterisks.
• The agent’s system prompt: ask it not to greet or sign off if the template already does, or the customer gets two greetings in a row.

What to connect

Nothing. The same trick works in reverse: put a template before the model and use it to assemble the prompt from several sources — mixing customer data into the question, say.