Behavior & prompt design
Designing the prompt that shapes Anna's behavior — goals, hard rules, and common pitfalls.
The prompt is the most important configuration choice. A good prompt makes Anna feel like a colleague; a bad one makes her sound like a chatbot.
Anatomy of a good prompt
- Identity: name, role, business name, tone.
- Goals (in priority order): what Anna should accomplish in priority order.
- Tools available: what she can do (book, lookup, transfer, take note).
- Hard rules: what she must never do (share private data without auth, speculate, give medical/legal advice).
- Style: turn length, formality, fallback phrases.
Common pitfalls
- Vague goals. "Help the caller" is too broad — be specific about which requests Anna handles end-to-end.
- No hard rules. Without explicit "never do X" lines, Anna will sometimes try to be helpful in ways that backfire (e.g. quoting prices she doesn't know).
- Long greetings. Anything over 12 words at the start makes the caller feel they're talking to a bot. Keep it tight.
Tested templates
Each vertical template ships with a battle-tested prompt — see Examples for ready-to-use starting points.
Advanced — creativity, precise mode, sliders
Under Callina → Assistant Configuration → Behavior:
Creativity (temperature)
0..1. Low = sticks to the prompt verbatim. High = freer phrasing. Default 0.5. Bump to 0.7 for sales/empathy contexts; pin to 0.2 for medical / legal scripts.
Speech speed
0.7..1.2 multiplier. ElevenLabs voice setting speed. Slower = clearer for older callers.
Sensitivity / interruptibility
0..1. Higher = caller can interrupt Anna more easily. Production tip: 0.6 for retail/customer-service, 0.4 for medical (where you want Anna to finish a structured question).
Max call duration / max silence wait
Hard caps in seconds. Cost protection — calls that hit the duration cap end cleanly with terminationReason='max_duration_reached'.
Precise information processing
Per-tenant list of fields where dictation accuracy is critical (email, phone, customer-number). Anna repeats the value letter-by-letter for confirmation before using it in any tool call. Configure under Callina → Assistant Configuration → Precise.