System Prompt
The full instruction set an agent runs on, its tokens and when to edit it.
Showing instructions for Xebit VoiceCloud.Showing instructions for Xebit VoiceKit, on your own infrastructure. Switch with the selector at the top. Updated 22 October 2026.
The system prompt is the complete instruction set an agent runs on. Nothing is added around it: what is in the system prompt is what the agent works from. A new agent starts with a working default, and most teams never need to touch it. Edit it when you want to change the structure itself: reorder sections, remove a block you do not use, or reword the standing rules.
System prompt or instructions?
The agent's instructions (its persona and task) and its greeting are separate fields. They flow into the system prompt through tokens: the instructions through {{other_details}} and the greeting through {{greeting}}. If you only want to change what the agent says or asks, edit the instructions, not the system prompt.
Tokens
The system prompt is plain text with {{token}} placeholders. On every call each token is replaced with a value or with a block of instructions built for that call. A token the platform does not recognise is removed and a warning is logged; it is never read to the agent as text.
Value tokens
| Token | Becomes |
|---|---|
{{agent_name}}, {{agent_number}}, {{agent_timezone}} | The agent's name, outbound number on your trunk and time zone |
{{agent_current_day}}, {{agent_current_date}}, {{agent_current_time}} | Today's day, date and the current time in the agent's time zone |
{{voice_style}} | The configured voice style |
{{primary_language}} | The language the call starts in |
{{supported_languages}} | The languages the agent may switch to |
{{greeting}}, {{ending_message}} | The greeting and closing message, with their own variables filled in |
{{other_details}} | The agent's instructions, with variables filled in |
{{customer_name}}, {{call_context}} | Context passed in when the call was started |
{{org_id}}, {{user_id}} | Identifiers for the call |
Caller details such as {{user_name}} and {{user_number}} | See Prompt Variables |
Section tokens
Each section token expands to a block of instructions for that call, or to nothing when the feature is not set up on the agent.
| Token | Expands to | Empty when |
|---|---|---|
{{current_datetime_section}} | The current date and time, so the agent can work out "next Friday" | Never |
{{appointment_section}} | How to book appointments and in which order to use the tools | Appointments are off |
{{callback_section}} | How and when to schedule a callback | Callbacks are off |
{{transfer_section}} | When and how to hand over to a person | Transfer is off |
{{callback_opening_section}} | An opening for a call that is itself a scheduled callback | The call is not a callback |
{{follow_up_section}} | The follow up actions available | None are configured |
{{precall_data_section}} | Data fetched by pre call workflows | There is no pre call data |
{{api_tools_section}} | The names of the during call workflow actions the agent may use | No during call workflows are enabled |
{{known_contact_section}} | A reminder not to ask again for details already on file | Name and email are both unknown |
{{call_metadata_section}} | The phone number, organisation and any custom data attached to the call | Never |
Keep the capability tokens. If you remove {{api_tools_section}} while the agent has during call workflows, it is never told which actions exist and cannot use them. The same applies to {{appointment_section}} and booking. The editor warns you when you save, but a call placed with the token missing will still go out without those abilities.
Standing rules
These tokens expand to fixed guidance every voice agent should carry. Keep them unless you mean to replace that guidance with your own.
| Token | Covers |
|---|---|
{{greeting_ending_conditions}} | Handling conditional wording in the greeting and closing |
{{ending_call_rules}} | When to end the call and how to say goodbye |
{{speech_format_rules}} | Speaking for the phone: no formatting, no reading out internal IDs |
{{voicemail_rules}} | Recognising an answering machine and ending the call |
{{filler_phrase_rules}} | Saying a short "one moment" before an action so there is no dead air |
The default template
A new agent's system prompt is laid out in this order: the current date and time; who the agent is and its call context; its role (name, greeting, closing, voice style, the language to start in and the languages it may switch to); the greeting and ending rules; then the detailed instructions, made up of the transfer, appointment, follow up, callback, pre call data and workflow action sections followed by your own instructions, the known contact note and the call metadata; and finally the speech, voicemail and filler phrase rules.
You can restore the default at any time from the agent's settings.
On Xebit VoiceCloud, Xebit maintains the default template and the standing rules.
On Xebit VoiceKit, the default template and standing rules come with your installed release. Read the release notes before an update in case they change.
Editing safely
- Change one thing at a time.
- Save, then make a test call that exercises the part you changed. See Agent Testing.
- Watch for prompt regressions after the change goes live. Observability flags them.
Next step
See Agent Prompting for writing the instructions themselves, and Workflows for the actions the workflow sections describe.