Instructions, languages and variables
Write a clear task prompt, set a first message and supply business variables.
Open the assistant's Assistant tab. Define the task and limits in plain language, then use the Visual preview to inspect the resulting instructions.
Start with a specific task
You help customers of {{organization_name}} prepare a support request.
Begin in English. Switch to Kiswahili when the caller requests it.
Ask one question at a time: their name, county, and the issue.
Answer service questions from the selected knowledge documents.
If a fact is missing, say so and ask a useful follow-up question.
Read back the request before proposing a submission action.
Confirm submission only after the tool returns a successful result.This is a starting prompt. Use Flow to enforce required intake, Knowledge for evidence, and tool approvals to authorize submission.
Set a short first message that explains what the assistant can help with. Choose assistant-first or user-first behaviour appropriate to the channel.
Use variables for business context
Write {{variable_name}} in the prompt or first message. In Advanced, define static values and an IANA timezone such as Africa/Nairobi.
| Built-in | Value |
|---|---|
assistant_name, organization_name | The assistant and workspace names. |
current_date, current_time, timezone | Date and time in the configured timezone. |
call_id | The current call identifier; previews use a preview identifier. |
Define up to 30 custom static variables. Undefined names block saving or starting the call. Values expand once and do not recursively expand more placeholders. Keep credentials in Integrations.
Choose the conversation language
Align the assistant language, recognition language and voice. Automatic recognition still needs a provider/model supporting the callers' languages. Test English, Kiswahili and switching with actual samples.
Validated-document answers preserve the configured or established language until the caller requests a switch. Knowledge can retrieve across languages, but extraction and answer quality still need scenario tests.
Prompt generation proposes text for review. Apply it deliberately, Save, and test before publishing.