Skip to main content
An assistant can be driven in two ways: a single system prompt or a visual flow. Both use the same voice engine; the difference is how much structure you impose on the conversation.

​
Single system prompt

The simplest setup: one prompt describes the assistant’s role, knowledge, and rules; one first message defines the greeting. The LLM handles the entire conversation freely within those instructions. Best for:
  • FAQ and reception assistants (“answer questions, take messages”)
  • Assistants whose job doesn’t branch into distinct phases
  • Fast prototyping — you can rewrite a prompt in seconds
Configuration:
  • Agent type — chosen in the Create assistant modal and later under Settings → General (Single prompt vs Conversational flow). Mode is stored as flow_json: null = prompt, object = flow.
  • System prompt — role, tone, rules, and facts. Use Choose template on the canvas (Prompt mode) or when creating the assistant. Applying a template copies into the prompt (and optional first message); it does not keep a live link to the template.
  • First message — the opening line, spoken when the call connects.
  • Greeting mode — agent speaks first (typical inbound) or user speaks first (the assistant waits; useful for outbound where the callee says “Hello?”).
  • Allow interruption (greeting_allow_interruptions) — optional; default off so the opening greeting plays to the end before barge-in. Separate from conversation-level Allow interruptions in Call settings. Ambient/background audio starts after the greeting finishes.
  • Audio greeting — optionally upload or record an audio file (mp3/wav/ogg/m4a, ≤ 5 MB) that is played verbatim at call start instead of the synthesized voice. It overrides the first-message text; the text is still used as the transcript and as the fallback for chat, voicemail, and when decode/playback fails at call start. Manage it via the POST/DELETE /assistants/{id}/greeting-audio endpoints or the Greeting card in the editor.
  • Assistant picture — optionally upload a PNG/JPEG/WebP image (≤ 1 MB; compressed server-side to ≤512×512 / ~400 KB) shown in the product UI instead of the default colored orb. Hover the avatar in the assistant header to upload or replace it. Manage it via POST/DELETE /assistants/{id}/avatar (not via PATCH).

​
Flow builder

The flow builder turns the call into a graph: multiple specialized agents, condition branches, HTTP tools, data-collection steps, transfers, and explicit endings. The engine hands the conversation from node to node. In the editor, switch to Flow (or create with Conversational flow). The base system prompt is edited under Settings → General → Advanced prompt (and collapsed on the canvas). Agent-node instructions are appended to that base — they do not replace it. Best for:
  • Calls with distinct phases (qualify → collect data → book → confirm)
  • Reliable data capture (names, emails, phone numbers with built-in validation)
  • Calls that must branch (“existing customer?” → different paths)
  • Transfers with rules (warm transfer to sales only after qualification)

​
Which should you choose?

A flow’s agent nodes can leave instructions empty — the assistant’s system prompt (Advanced prompt) is used as the base. When an agent node has its own text, that text is appended to the system prompt, not substituted. Start with a prompt and add flow structure without duplicating the whole persona.

​
Settings that apply either way

Regardless of mode, every assistant has: engine mode (pipeline / realtime / half-cascade), model and voice selection, knowledge base, conversation-quality settings, recording and consent options, max call duration, idle timeout, and a webhook URL for call results. iOS / Android Call Screen Handling (Settings → Conversation, or the Pre-Call node in the flow builder) lets the agent answer screening services (“Who’s calling?”) with a name, company, and reason, then wait for a human. Stored on pre_call and mirrored into flow_json.pre_call when a flow exists. On outbound calls the worker runs answering-machine detection — screening prompts classify as machine-ivr, then the agent speaks the configured identity reply and waits for a real person before the normal greeting.