Instructions and customisation

Custom instructions tell Tori how you want it to handle customer questions. You can use them to change how it speaks, how it researches, when it asks your team for help, and other parts of its behaviour.

Tori already has instructions for doing these things. You only need to explain what you want it to do differently. Instructions you don't change continue to apply.

Start with the defaults

You don't need custom instructions to start using Tori. Give it your website and try a question that a customer might ask. Choose a question that needs information from your website so you can check that Tori can read it.

As you use Tori, you may find something you want to change. Add an instruction for that change rather than writing a new description of its entire job.

The behaviour reference describes Tori's default instructions. It also gives examples of how to change them. Those examples aren't a complete list of what you can customise.

Writing instructions

Write instructions in plain language. Say what you want Tori to do and, where it helps, when it should do it.

For example, to change what Tori calls itself:

Call yourself Tori.

To change how it handles languages your team doesn't support:

Our team only provides support in English. If the customer asks a question in another language, first apologise in their language. Then repeat their question in English and answer in English.

To change when it asks for human help:

For billing questions, escalate immediately. Don't try to answer because you can't access our billing system.

An instruction should not need repeated warnings or several paragraphs saying the same thing. A long set of instructions can make sense if you have many different requirements, but you shouldn't need a long explanation of each one.

If Tori repeatedly ignores a clear instruction, contact us. You don't need to keep rewriting it before asking for help.

Instructions and sources

Use sources for information Tori should read when answering questions. Use custom instructions for how it should handle those questions.

For example, your refund policy could be a source. An instruction could tell Tori to explain that policy, then escalate if the customer wants someone to approve an exception.

This isn't a strict division. Custom instructions can also contain facts or correct a common mistake. You can give a short correction directly in an instruction, or put a longer explanation in a source.

A source description has a more specific purpose. It helps Tori decide when to read that source and how to use the information in it.

What instructions can change

Instructions guide Tori's decisions. They aren't code, so they don't guarantee that every reply will follow a rule exactly.

They also don't give Tori new access or tools. Telling Tori to explain your refund policy is different from telling it to issue a refund. Tori can explain the policy, but it cannot perform an action in your billing system.

Only include information you would be willing to share with customers. A customer may persuade Tori to repeat its instructions, even if you've told it not to.

Testing changes

The test chat uses the same sources and instructions as the agent on your website. Saving an instruction changes the agent, not just the test chat. There is no separate publish step for instructions.

Start a fresh conversation when you test a change. Otherwise, earlier messages may give Tori information it wouldn't have in a new customer conversation. See Conversations and context.

To try changes without affecting customers, create a separate agent with the sources and instructions you want to test. After testing, make the changes to the agent your customers use.

Separate agents

An agent has its own sources and custom instructions. The chat on your website is a way for customers to talk to that agent.

You can use the same agent in more than one place. Changing its instructions changes how it handles conversations wherever you use it.

Create separate agents when you want different sources or instructions. For example, you might have separate agents for two products, different audiences, or testing. Changes to one agent's instructions don't change another agent's instructions.