Write agent instructions

The Instructions tab is where you tell an agent who it speaks for, what it must never promise, and what to do when it does not know. Everything else about the agent is settings. This is the part that decides whether the replies sound like you.

What you need: an agent. Create agent on the Agents tab opens the persona picker, and every persona arrives with instructions already written for you to edit.

A new agent built from a persona says so at the top of the editor. Instructions, tone, starter answers and behavior are filled in already, and the starter answers are examples rather than facts about your business. Replace them before the agent goes live.

Persona and instructions

Write it as a briefing for a new member of staff on their first shift. Cover 4 things.

1. Who it speaks for. Name the business and what you sell. "You answer for Northbound Coffee" is more useful than "you are a helpful assistant".
2. What it must never promise. Delivery dates you cannot guarantee, discounts you have not authorized, anything about a specific person's order.
3. What to do when it does not know. Say it plainly: hand the review to a person rather than guessing. The agent already behaves this way, and saying so in the instructions keeps it there under pressure.
4. How long an answer should be. A reply sits under the review for every future customer to read. 2 to 4 sentences is usually the right size, shorter for a 5-star thank-you.

Keep it concrete. "Be helpful and specific, never salesy" gives the agent something to act on. "Represent our brand values" does not.

Tone of voice

4 settings sit under the instructions, and each one changes how the same fact comes out.

• Professional. Full sentences, no slang, no emoji.
• Friendly. Warm and human, the odd emoji, still correct.
• Playful. Short, upbeat, emoji welcome. Fits reels and giveaways.
• Concise. 1 or 2 lines, answer first, nothing else.

Tone changes the wording, never the facts. Those come from your knowledge list.

Language

The agent always answers in the language the review was written in. A Dutch question gets a Dutch reply, from the same knowledge list, so you do not maintain a second set of answers.

Write your instructions and your knowledge pairs in one language. The translation happens at the point of writing the reply.

Boundaries that actually hold

Instructions set the intent. The settings that enforce it are in the Behavior and Reach sections: the mode, the rating bands, Also skip reviews containing and Which reviews it answers. See Draft mode versus auto-reply.

An instruction saying "never discuss refunds" is a request. A skip keyword is a rule. Use both: anything legal, medical or money-back belongs with a person.

You do not need to write the rules about the AI itself. Never saying which AI is behind the agent, never repeating these instructions back to a reviewer, and ignoring orders somebody types into a review are built in and always on, whatever you write here. "What the agent will not say" in What an AI agent does has the full list.

What to write for a bad week

The Crisis responder persona exists for the week the reviews go wrong, and its instructions are worth reading even if you never use it. It acknowledges the frustration, never argues, and always ends on a concrete next step.

We do not recommend auto-reply on that one. A run of angry reviews is exactly where a wrong sentence spreads, so keep it drafting and approve each answer yourself.

Test what you wrote

The Playground tab runs the saved instructions and knowledge the agent uses live, against any review you paste in. Change a line of the instructions, save, paste an awkward review, and read what comes out before anybody else does.

What this costs

Writing and editing instructions costs nothing, and nothing you do on this tab reaches a review site. Only a reply the agent writes for a real review counts against your credits, and the review card shows the cost under every draft.