Write prompts

A prompt for a voice agent is not a chatbot prompt with a phone attached. The caller can’t scroll, can’t skim, and won’t wait — every reply is heard once, in real time, by someone who may be driving. The prompts that work are short, specific about the goal, and silent about the words.

Two prompts, two jobs

The split matters because of what each one is paid for. The System Prompt is resent on every turn of the whole call, so it should be a few lines. Agent Prompts are only active on their own step, so they can afford detail — but the same rule applies: each one describes one job.

What a good prompt looks like

You are Mia, the phone assistant for Northwind Dental in Seattle.
This is a live phone call: keep replies to one or two short sentences and never read lists aloud.
You are not a clinician — never give medical advice; offer to take a message instead.
If the caller is upset, stay calm and offer a human callback.

Goal: book, move or cancel a hygiene appointment.
Collect, one at a time: full name, date of birth, preferred day and time.
Read the collected details back and get an explicit yes before calling book_appointment.
If the requested slot is unavailable, offer the two nearest alternatives from the function result.
When the booking is confirmed, thank them and end the call.

What these do right, and why:

And what to leave out:

• Scripts. “Say: ‘Hello, thank you for calling…’” produces robotic delivery and breaks the moment the caller interrupts. Describe the intent; let the model phrase it.
• Formatting. Bullets, bold, headings — the caller hears none of it and the model may read the punctuation.
• Everything the agent doesn’t need on this step. A booking agent doesn’t need the refund policy. Long prompts cost tokens on every turn and dilute the instructions that matter.
• Contradictions between the two prompts. If the System Prompt says “be brief” and an Agent Prompt says “explain the full policy”, the model picks one — not necessarily yours.

Emotion and delivery with ElevenLabs v3

Most voices say exactly what the model writes, in one register. ElevenLabs v3 (eleven_v3, eleven_v3_conversational) also reads inline audio tags — [warmly], [laughs], [whispers], [hesitates], [excited] — and performs them. On a v3 voice, your prompt can direct delivery as well as words:

Your speech is rendered by an expressive voice that understands bracketed delivery tags.
Use at most one tag per reply, only where it changes meaning: [warmly] when greeting,
[apologetic] when you can't help, [laughs] only if the caller made a joke. Never tag routine sentences.

Rules that follow from how v3 works on Talkif:

• Only on v3 models. On every other voice a tag is read aloud as text. If you switch the TTS node, remove the tag instruction from the prompt in the same publish.
• Tags appear in the transcript as written, so a reader sees [warmly] Good morning. Acceptable for most teams; if transcripts feed a downstream system, strip [...] there.
• Fewer voice sliders. On v3 only stability applies; similarity, style and speaker boost are ignored.
• Keep tags rare. One per reply at most — over-tagged speech sounds theatrical on a phone.

ElevenLabs’ own guide to v3 tags and phrasing: Prompting Eleven v3.

Dynamic prompts

Both prompts are templates, resolved when the call starts. Anything Talkif knows about the call can be inserted with {{…}}; type {{ in the editor for suggestions.

You are calling {{default contact.firstName "the customer"}}
{{#ifExists contact.company}}from {{contact.company}}{{/ifExists}}
about their appointment. Today is {{formatDate system.date "long"}}.
{{#if (eq call.direction "inbound")}}Thank them for calling.{{else}}Confirm it's a good time to talk.{{/if}}
{{#ifExists contact.notes}}Background from our records: {{contact.notes}}{{/ifExists}}

Variables

A variable with no value renders as empty text — never an error, never the literal {{…}}. That’s why default and #ifExists are worth using on every contact field: not every caller is in your contacts.

Helpers

Truthiness is voice-friendly: empty strings, "none", "null", empty lists and empty objects all count as false. Output is plain text — nothing is HTML-escaped.

The Prompts workspace

Flow Builder → Prompts is a place to write and keep prompts outside any single flow: company-wide tone guidelines, a persona you reuse across lines, a booking step several flows share. It has the same editor with {{ autocompletion, plus two things the node editor doesn’t:

• Preview with contact data — pick a real contact and see the prompt exactly as the call would render it, variables filled in.
• AI generation — describe what you want and iterate in chat; the generator knows your variables and helpers and produces a template that’s ready to use.

Prompts here are copied into agents, not linked: paste the text into the node. That keeps each published version self-contained, and it means editing a workspace prompt never silently changes a live flow.

Next