Configuring an assistant

Configuring an assistant is mostly about one thing — writing good instructions — and a few supporting choices. This page walks through each setting, what it controls, and how to choose well.

Instructions: the persona

The instructions are the heart of an assistant. They're a written brief — its system prompt — telling the AI who it is, how to behave, what it handles, and what it should never do. Everything else is supporting cast; this is the lead.

Good instructions tend to cover:

• Who it is — "You're the support assistant for Acme, a Turkish home-goods store."
• How it talks — warm, concise, professional; whatever fits your brand.
• What it does — answer product and order questions, help with returns.
• Its limits — never promise discounts, never guess at shipping dates, hand off for refunds over a certain amount.

Model tier: how hard it thinks

The model tier decides how much brainpower the assistant uses. There are three, and they trade speed and cost against raw capability:

• Fast — quick and economical. Great for high-volume, straightforward chat where speed matters more than deep reasoning.
• Balanced — the sensible default. A strong mix of quality and speed that suits most assistants.
• Smartest — the most capable, for complex reasoning and the trickiest conversations, at a higher cost per reply.

A good instinct: start on Balanced, and only move to Smartest if you see the assistant struggling with genuinely hard reasoning, or drop to Fast if a simple assistant is handling huge volume.

Creativity (temperature)

Creativity controls how predictable the assistant is. Lower means more focused and consistent; higher means more varied and exploratory.

For support and sales, lower is almost always better — you want reliable, on-message answers, not improvisation. The default leans factual for exactly this reason. Save the higher settings for genuinely creative tasks, which are rare in customer conversations.

Response length

You can cap how long a reply can get. Most support answers are short, so the default is comfortable. Raise it if your assistant needs to produce longer, more detailed responses; keep it modest if you want crisp, to-the-point replies.

Knowledge and tools

An assistant becomes genuinely useful when it can reach your knowledge and take actions. Through the assistant — and especially the Agentic AI capability — it can search your collections, remember details, look up customer context, and more. What it can reach is part of its configuration, so you can give a focused assistant exactly the tools its job needs and nothing more.

Guardrails

Finally, you can attach guardrails to an assistant — safety and compliance rules that check messages on the way in and on the way out. This is how you keep an assistant on-topic, protect sensitive data, and enforce the rules your business cares about. Guardrails are important enough to have their own section; start at Guardrails overview.

A sensible starting point

If you're configuring your first assistant, this is a solid baseline:

1. Write clear, specific instructions — the single highest-impact thing.
2. Leave the tier on Balanced and creativity low.
3. Connect the knowledge it needs.
4. Attach a PII guardrail so sensitive data is handled safely from day one.

Then watch real conversations and adjust. The best assistants are tuned by observation, not guessed at up front.

→ Next: Agentic AI — how an assistant actually gets work done.