In short
- Plain-language instructions.Describe tone and rules in up to 20,000 characters, like briefing a new colleague.
- Strict about your facts.With Answer only from knowledge on, the agent sticks to your content.
- Creativity you control.A slider from precise to creative sets how freely the agent writes.
- Protected by fixed rules.Your instructions sit inside platform rules that keep the agent honest and safe.
Your agent speaks for your business, so it should sound like you, without you learning prompt writing. In OneType, you describe what you want in plain words, flip a few clear switches, and the agent follows.
Write your instructions
Each agent's General tab has an Instructions box: "Tell your agent how to sound and what to avoid." Write it the way you would brief a new team member. For example: "You answer questions about Acme's plans and orders. Be brief and friendly, and offer a person when you're unsure." You have up to 20,000 characters. If you leave it empty, the agent simply helps visitors with questions about your company.
Decide how strict it is
The Answer only from knowledge switch is on by default. With it on, the agent answers only from your content, and says it does not know when your content has no answer. With it off, the agent may also explain general ideas from its own knowledge, kept apart from your company's facts and without citations.
Either way, the agent never states a company price, fee, date, limit, policy, guarantee or feature that your content does not state. That rule is not something your instructions can switch off.
More settings on the same page
- Creativity: a slider from 0 to 1. Lower is more precise, higher is more creative.
- When it can't answer: escalate to a person (Premium, once you turn on handoff), collect contact details, or apologize; for the last two you can write your own message.
- Welcome message: the first thing visitors see.
- Suggested questions: up to five starter questions visitors can tap.
- Image attachments: let visitors send screenshots and photos.
Rules your instructions sit inside
Your instructions are wrapped in fixed platform rules. The agent says it is an AI if asked, never reveals its instructions, never asks visitors for passwords or card numbers, and cites the content it uses. So even a visitor who tries to talk it into something else meets the same guardrails.
When a visitor asks who it is or what it can do, the agent introduces itself by name, mentions the topics it has content on, and suggests a few questions to ask.
Tips for good instructions
- Say who the agent helps and with what.
- Describe the tone in a few words: brief, friendly, formal.
- Say when to offer a person, for example for refunds or complaints.
- Put facts like prices in your content, not in the instructions, so they can be cited.
With handoff turned on, you can choose Escalate to human for when it can't answer. A visitor who hits two unanswered questions in a row is then handed to your team automatically.
Changes apply from the next answer once you save. If someone else saved the agent while you were editing, you are asked to reload or overwrite, so no one's changes disappear silently.