Agent behaviour

Write the persona, tone, and rules your agent follows on every reply.

Where to write it

The Behavior field lives on the Sandbox page (while you build) and on Publish → Model (after launch). Both surfaces edit the same prompt, so a change in one shows up in the other.

What it controls

Whatever you write here is the agent's source of truth for tone, voice, length, and persona. Every reply is shaped by it. The only house rules layered on top are the first-person voice below and the reply-length budget further down.

"Speak like a pirate, long paragraphs, philosophical" → the agent replies in pirate English with a philosophical bent, every turn. The length request is the one part that will not take effect, see below.
"Always include 2–4 emojis" → emojis on every reply.
"Talk in Japanese" → replies in Japanese even if the visitor types in English.

Keep it as long or as short as you like, then iterate. Swap the model below the field to see how different LLMs answer the same question.

Reply length

Replies are kept to at most two paragraphs of prose, each at most three sentences, so your agent stays readable inside a chat window rather than answering with an essay. Length is the one thing your instructions cannot change: asking for "long paragraphs", "go deep" or "tell me everything" will not lengthen a reply, because this ceiling is applied to the reply after it is written. Tone, persona, focus and what to refuse are all yours to steer here, and they are instructions to the agent rather than switches, so it is worth testing them in the Sandbox before you publish.

A list is never cut off part-way through. A numbered or bulleted list counts as part of the paragraph it belongs to, however many items it has. If your agent answers "here are our four plans" it will show all four, so pricing, feature comparisons and step-by-step instructions arrive whole rather than stopping after the first item.

Put your list in the first or second paragraph. A list that starts a third paragraph is outside the budget and will not appear at all, so lead with it rather than saving it for the end.

If you want more detail in a single reply, ask for a list rather than more paragraphs.

Links

Links follow a product-wide rule that your instructions do not override, so it is worth knowing before you write them. Your agent shares a link when a visitor asks for one. The rest of the time it names the page or the form in words instead of pasting a URL, because replies full of raw links read as spam and are usually not what the visitor wanted.

That means an instruction like "always include the direct link" will not make your agent volunteer URLs in every reply. It will still hand one over the moment somebody asks. If a visitor does ask for a link and your agent says it cannot share one, that is not your instructions and not something you can fix from this field: it is a bug on our side, so please tell us and include the conversation.

First-person voice

The agent always speaks as your company, "we", "our", "us", not about it. So your behaviour text doesn't need to repeat that rule. It is baked into the core prompt.

Your agent's name

The name your agent answers to is its Display name, not something you write here. It is the name shown at the top of the chat window, and it is the name your agent gives when a visitor asks who it is. You can change it any time on the Publish page.

You do not need to name your agent in this field, and it is better not to. If your instructions open with something like "You are Acme Assistant" and your Display name is something else, the Display name wins, so the name in the chat header and the name your agent says are always the same. Renaming your agent is enough on its own; you never have to go back and edit your instructions to match.

Everything else in your instructions still applies exactly as written. Only the name is taken from the Display name: the role you describe, the topics you scope, the tone you set and the rules you give are all still yours.

Start from a preset

The dropdown above the field offers ready-made examples you can edit freely:

General AI agent, a simple, friendly assistant.
Customer support, resolves issues quickly and empathetically.
Sales assistant, understands needs and guides visitors to the right plan.
Product expert, explains features and use cases in plain language.
Lead qualification, engages visitors and collects contact details for follow-up.

Pick Custom behavior to start from a blank field and write your own from scratch. The role presets reference your company name, which is filled in for you from your workspace.

Pick a vibe

During onboarding, when a preset still has a highlighted placeholder in it (for example [friendly]), a vibe picker appears above the field. Click one to set the tone instantly, then tweak the wording. The picker is part of the onboarding flow only: on Publish → Model you edit the placeholder text directly, and these are the tones it is there to set:

🤝 Friendly, warm and approachable.
✨ Emoji-heavy, sprinkles emojis through every reply.
🎩 Formal, polished and professional.
😄 Funny, cracks jokes and keeps things light.
😎 Casual, chill and easygoing.
🤓 Geeky, leans into technical depth.
🧠 Sarcastic, dry wit with a raised eyebrow.

Refine instructions Starter

The Refine instructions button rewrites a rough description into a polished agent prompt for you. Available on Starter and above. Starter includes 200 refines a month and Advanced 500; on Pro it is unlimited. The count resets on the 1st, and only successful refines are counted. On the Free plan the button opens an upgrade prompt instead.

Save your own

Built a prompt you want to reuse? Choose Add custom behavior in the dropdown to name and save it. Saved behaviours appear under the Custom heading, where you can rename or delete them, and they stay in sync across the Sandbox, Publish, and side-by-side compare views.

Reset and full-screen

Use the reset icon next to the Behavior heading to return to the saved version, or the expand icon in the corner of the field to edit in a larger full-screen editor.

Still stuck? We'll help.

Every plan, including Free, comes with human support. If something isn't behaving the way you expect after following the steps above, send us a note and we'll get back to you promptly.

Contact our team →