What Answer instructions are
Answer instructions — what other tools call a system prompt — are a set of rules and background that go out with every message the AI writes. They shape how it introduces itself, how formal or casual it sounds, what it will and will not discuss, and how it handles awkward cases.
You will find the field at Brand → AI agent → Answers, in the Writing style group, labeled Answer instructions. The description under it reads: "Sent with every answer: rules, tone details, or facts Yaplet should always know." The same text appears as Instructions on Brand → Playground, which is the easiest place to try a change, because its test chat already answers with instructions you have not saved yet. Save with Save changes in the title bar. The instructions belong to that one agent, so a second brand has its own set and nothing is shared between them.
What to write in them
Think of it as the briefing you would give a new support hire on their first morning. Cover:
- Role: what the agent is and who it is for ("You are Yaplet, a support agent for Acme Corp, helping customers with billing and product questions.")
- Tone: formal, friendly, concise, warm — pick one or two words and stay with them.
- Topics to avoid: competitors, legal advice, medical advice, whatever is off-limits for you.
- Things to always do: sign off with a contact address, use the visitor's name, point at the knowledge base first.
- What to say when it does not know: the behavior itself is a separate setting in the Human hand-off section (see below), but you can reinforce the wording here.
- General knowledge: Yaplet states facts — about your company, your products, prices, policies and delivery, and about the visitor's topic — only from this brand's own data and articles, and never adds anything from its own general knowledge. Asked what something is or how it works, it answers only from your data, or says it has no more detail on that. This is the same for every brand, and these instructions cannot switch it off.
Tone settings beyond the instructions
One setting in the same Answers section changes the voice without you writing a word: Tone. Friendly, the default, adds an emoji now and then to keep the chat light; Formal gives the same answers, never with emojis.
Your company name is no longer set here. It is a brand setting, at Brand settings → General, alongside your website and a description of the business — and Fill from website can write all three for you from your website address. The AI reads them, so it is worth filling in properly even though it is not on this screen.
When Yaplet doesn't know the answer
In the Human hand-off section, When Yaplet has no answer decides what happens when nothing relevant turns up in this brand's knowledge. There are two choices:
- Politely decline: the agent tells the visitor it does not have information about that topic. It does not guess and does not make things up. Safest for most setups.
- Connect to a person: if someone from your team is online, the visitor is connected to them right away, with no question asked first. If nobody is available, it still hands over: the agent tells the visitor that nobody from the team is available right now, and the website chat shows the card for leaving a message for the team.
The second option is grayed out until the Visitor asks Yaplet for a person switch, in the Ways to ask for a person group above it, is on, with the hint "Switch on “Visitor asks Yaplet for a person” to use this." There is no third "answer anyway from general knowledge" option — that choice does not exist in the product.
Tell visitors how to sign in
Also in the Answers section is the group How visitors get identified, with the field Message to unidentified visitors. API tools and workflows can be switched to Only for identified visitors, so they work only for customers your website signed in through Yaplet.identify (see Identify your logged-in users with custom data). When a visitor who is not identified asks for one of them, the AI does not run it. It passes this text on instead, in its own words and in the visitor's language.
Write where and how your customers sign in, for example "Log in at example.com/account, then open the chat on any page and ask again." It is plain text, at most 1,000 characters. Left empty, visitors are told the standard sentence: "To use this, please sign in to your account on our website and open the chat from there, then ask again." Next to the switch on an API tool and on a workflow's Starts when card, a panel shows the text visitors get right now, with an Edit this text link that brings you back here. A Call an API step has its own Only for identified visitors switch; with it on, a visitor who is not identified does not get this text but takes the step's Failed exit.
Tips for effective instructions
- Be specific, not vague. "Be friendly" is hard to act on. "Use first names when the visitor introduces themselves, never say 'ticketing system', and keep replies under four sentences" is actionable.
- List what NOT to do. Restrictions usually land harder than positive instructions.
- Keep it short. A few tight paragraphs beat a page of prose — these instructions go out with every single message, and a long one dilutes the rules that matter most.
- Test after every change in the Playground (Brand → Playground), where the test chat already uses instructions you have not saved yet.
Example instructions
You are Yaplet, a support agent for Acme Corp. Your job is to help customers with questions about our project management software.
Tone: Friendly but professional. Use first names when offered. Keep replies concise — 2–3 sentences max unless a step-by-step answer is needed.
Always: Suggest the knowledge base first. If you can't find the answer, say so clearly and offer to connect the visitor with a human agent.
Never: Discuss competitors. Give pricing advice beyond what's in the knowledge base. Make promises about features that aren't documented.
If asked about your technical nature, say: "I'm Yaplet, Acme's AI support assistant."
Next step
Try your wording before real visitors see it: Test your chatbots before going live. To control when the agent escalates to a person, see Control when Yaplet hands off to a human.