Instructions & persona
Write the persona and rules that shape your agent — the editor, ready-made templates, variables, insert-blocks, and how instructions reach the AI.
The Instructions page is where you define who your agent is and how it behaves. This persona is the agent's operating manual: its role, tone, the flow it follows, its hard rules, and when it should escalate. Getting it right matters more than almost anything else you configure.
The editor
Instructions are written in a rich text editor that supports headings, bold, italic, inline code, bulleted and numbered lists, and links. A side panel offers tips, variables you can insert, and a button to apply a template.
The editor shows a live character count and an approximate token count.
Instructions are capped at 10,000 characters. A soft warning appears at 8,000 characters, and Save is disabled above 10,000. Beyond a couple of thousand tokens the AI actually follows instructions less reliably — if you need to convey a lot of facts, put them in Data sources, not the persona.
Templates
Don't start from a blank page. Click Apply template to drop in a proven persona, then edit it. Templates are grouped by category:
Prop
Type
If the editor already has content, applying a template asks you to confirm before replacing it.
Variables
Insert variables that resolve automatically on every conversation, so you never hard-code details that might change:
Prop
Type
Variables are stored exactly as written and filled in at reply time, so renaming the agent instantly updates the persona without re-saving. The editor renders them as small chips.
Insert-blocks
Type / in the editor to insert a ready-made block of structure. These scaffolds
help you write a persona the AI follows well:
- Persona — a role + brand opener using
{{agentName}}/{{businessName}}. - Conversation flow — collect one field per turn, in order.
- Reply cadence — an acknowledge → reason → ask rhythm.
- Hard rules — things the agent must never do.
- When to escalate — triggers for handing off to a human.
- Example exchange — a worked
Visitor:/You:example to imitate. - Off-topic deflection — steering the conversation back on track.
How instructions reach the AI
Your instructions are given to the model as its authoritative operating manual. The agent follows the flow, tone, and rules you write. A few things are worth knowing:
- Safety always wins. Platform guardrails — staying on scope, grounding answers in your knowledge, resisting prompt-injection from untrusted content — take priority over the persona if they ever conflict.
- Your flow beats the defaults. If your persona says the first question should be "Are you looking to buy or rent?", the agent asks that instead of a generic "How can I help?".
- First-message behaviour. When a visitor just says "hi", the agent opens
your flow rather than a canned greeting. That's why good templates start with a
clear first message and include a
Visitor: "hi"example — the agent imitates what it sees.
Write the persona as a manual, not a wish list. State the role, the exact first message, the step-by-step flow, the hard rules, and one or two example exchanges. The clearer the structure, the more reliably the agent follows it.
What belongs elsewhere
The Instructions page is only for behaviour. Put other things where they belong:
- Facts and FAQs → Data sources
- Tools and integrations → Actions
- Greeting, colors, starters → Playground