> ## Documentation Index > Fetch the complete documentation index at: https://wiredesk.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # Writing instructions > What the agent's instructions control, how they combine with knowledge and WireDesk's own rules, and how to write ones that work. The instructions tell the agent about your business and how to behave: who it's talking to, what it helps with, its tone, and your house rules. They apply on every channel, unless you give a channel its [own wording](/docs/persona/channel-wording). Edit them at **Agents → your agent → How it talks**, in the box labelled **About your business & how to behave**. Then click **Save changes**. The same field is also on the **Agent** tab of the **Playground**, so you can edit and test side by side. | Limit | Value | | ------------ | ----------------------------------------------------- | | Length | Up to 4,000 characters | | Required | Yes. An agent can't be saved with empty instructions. | | Who can edit | Owner and Admin roles | ## Instructions are for behaviour, facts are for knowledge This is the single most important thing on this page. Instructions are **how the agent behaves**. [Knowledge](/docs/knowledge/overview) is **what it believes to be true**. Your agent treats its knowledge as more reliable than its instructions. If the two disagree on a price, an hour, an address or a policy, it gives the knowledge answer and ignores the instructions. If the knowledge says nothing on the point, it states what the instructions say, as fact. So a price typed into the instructions is quoted with full confidence whenever your knowledge doesn't mention it. It has no source you can update, and re-reading your website never corrects it. Put prices, hours, addresses and policies in knowledge, usually as a [note](/docs/knowledge/notes), and keep the instructions for rules about conduct. The agent's setup checklist catches the common case: **Prices and hours in the knowledge base** appears on the agent's **Overview** when the instructions contain a price or a time of day. ## How the agent's instructions are put together Your instructions are one layer of what the agent is told at the start of each conversation, in this order: 1. **Who it is.** It's told its [name](/docs/persona/name-and-greeting), and that it answers on behalf of your business. 2. **WireDesk's rules for the channel.** How to write for a phone call, a chat window or an email, plus honesty rules. You can't edit these. 3. **Your instructions.** Or that channel's own wording, if it has some. 4. **How to use its tools.** Knowledge search, notes, taking messages, handover, and any [actions](/docs/actions/overview) you've set up. Because of layer 2, you don't need to write rules about format or honesty. The agent already: * **Doesn't guess.** If it doesn't know something and nothing it looked up says, it says so plainly. It never invents prices, policies, availability or facts about your business. * **Doesn't describe how it works.** It won't say "according to my information" or "my instructions say", and won't name its tools. It answers the way a member of staff would. * **Writes for the channel.** On the phone: short, spoken, with no lists or symbols. In chat: a sentence or two, bold for key details, lists for three or more options, links only to addresses it was given. By email: a complete plain-text reply with a greeting, every part of the question answered, and a sign-off with the business name. * **Offers to take a message** when someone needs something it can't do. If there's no way to reach a person on that channel, it says so instead of ignoring the request. ## Draft instructions If you're starting from scratch, WireDesk can write a first draft for you. Use the menu next to the field: **Answer questions about the business**, **Handle support requests**, **Book appointments and capture leads**, or **Something else — I'll describe it**. Type a few notes about your business in the box. The agent's name alone is enough to start, but more detail gives a better draft. The draft replaces whatever is in the box. It has four headed sections: **Business Context**, **Role**, **Style** and **Guardrails**, and is under 350 words. Read it all. Remove any facts that should be in knowledge instead, correct anything wrong, then click **Save changes**. Nothing is saved until you do. The draft only uses facts from your notes, but check them anyway. If drafting fails, you'll see "drafting failed — try again". ## Write instructions that work Keep them short. Long instructions water down the rules that matter, and tempt you into writing facts where rules belong. The best agents usually have short instructions and plenty of knowledge. * Say who the customers are and what the agent is for. * Give house rules: "Never promise a delivery date." * Say when to hand over: "Anything about a billing dispute, offer to pass it to the team." * Set the tone: "Friendly and plain. No jargon." * Say what to ask for: "For an order question, ask for the order number first." * Paste your price list or opening hours. * Tell it to always have an answer. It will make one up. * Repeat the format rules WireDesk already applies. * Describe tools it doesn't have ("transfer them to sales"). * Write different rules for each channel here. Use [wording per channel](/docs/persona/channel-wording). ### Example A weak version, mixing facts in with rules: ```text theme={null} You are the assistant for Northwind Dental. Cleanings are $120 and whitening is $350. We're open 8-6 Monday to Friday. Always answer every question. Be super friendly!! ``` A stronger version, with rules only. Prices and hours are in knowledge: ```text theme={null} Northwind Dental is a family dental practice. Most people contacting us are existing patients booking or moving appointments, or new patients asking whether we can see them. - Be warm and plain. Short answers. - For anything about a specific treatment plan or a bill, don't discuss details. Offer to have the practice team follow up, and take their name and number. - If someone describes severe pain, swelling or bleeding, tell them to call the practice directly or seek urgent care. - Never give medical advice or say whether a treatment is right for someone. ``` ## What instructions can't do * **Give the agent new abilities.** Booking meetings, looking things up in your systems, transferring calls and letting a person join a chat are [actions](/docs/actions/overview). Writing "book them in" doesn't let it book anything. * **Override your knowledge.** If knowledge and instructions disagree on a specific fact, the knowledge answer wins. * **Switch off WireDesk's rules.** Honesty, channel style and not describing how it works always apply, including on a channel with its own wording. * **Change the opening line.** The first thing said is the [greeting](/docs/persona/name-and-greeting), played exactly as written. * **Add buttons or chips.** Suggested questions in the chat widget are set on the **Content** tab of the **Playground**, because they're buttons, not text. * **Change the voice or the AI notice.** Those are under **Channels → Phone**. ## Test your changes Save, then ask the agent something in the **Playground**. With the chat channel on, replies there come from the saved agent, so unsaved edits aren't used. Test chats appear in [Conversations](/docs/workspace/conversations). ## Related * [Wording per channel](/docs/persona/channel-wording) * [Fix a wrong answer](/docs/knowledge/fix-a-wrong-answer) * [How answers work](/docs/concepts/how-answers-work) * [Test your agent](/docs/agents/test-your-agent)