The AI Cold-Caller Prompt Vault

5 production outbound voice agents that book real meetings.

These are not templates written for a lead magnet. They are the actual system prompts, first messages, tool definitions, and evaluation criteria from five outbound voice agents we built and ran for real campaigns, sanitized for reuse: names, IDs, webhooks, prices, and client claims are swapped for {{placeholders}}, and everything else ships as it ran. Paste a pack into ElevenLabs Conversational AI and you have a working outbound agent. The prompts are plain text, so Vapi and Retell work too.

What's in the vault

Pack	Market	Language	Close
`dubai-brand-outbound/`	Dubai brand owners	EN	Books a 10-min Vision Call (Calendly tools)
`us-brand-outbound/`	US brand owners	EN	Books a 10-min Vision Call (Calendly tools)
`us-towing-dispatch/`	US towing / roadside owner-operators	EN	Books a 10-min phone audit (webhook tools, optional two-path recording variant)
`restaurant-outbound-et/`	Estonian restaurants	ET	Two-path: personalized demo SMS or booked call
`website-redesign-outbound-uk/`	UK local businesses	EN	Double-qualified email capture + next-day founder callback

Every pack has a PACK.html with deploy steps, the market-register reasoning, and a customize-first list. Packs include first-message variants (voicemail scripts where the campaign used them), tool definitions with documented request schemas, and paste-ready evaluation criteria where the agent had them.

Quick start (10-15 minutes per pack, no code)

Pick the pack closest to your market and read its PACK.html.
elevenlabs.io > Conversational AI > Create Agent.
Paste the first message and the system prompt. Fill the {{placeholders}}, or define them as ElevenLabs dynamic variables. Name your agent in the first line.
Set the voice and LLM per the pack's settings (each PACK.html lists what the source deployment ran).
Add the tools. The two brand packs use ElevenLabs' native Calendly integration, the towing and restaurant packs use plain POST webhooks (n8n, Make.com, or any HTTP endpoint), and the UK pack needs none.
Import a Twilio number on the Phone numbers tab and test-call your own phone before any real lead.

Adapting to Vapi or Retell: the prompts port as-is. Tools need re-mapping to each platform's custom functions; every PACK.html includes the mapping. Keep the tool names and the "call IMMEDIATELY after X" phrasing in tool descriptions, that timing language is what makes the model fire them at the right moment.

The methodology on one page

Distilled from Andres Contreras's public material plus our own production calls. Every pack runs some version of this spine:

Radical-honesty open. "Quick honesty, this is a cold call. Got 30 seconds? Then you decide." Disclose and offer the exit before any value statement. It is counterintuitive and it survives more openers than any disguise.
Permission before pitch. The agent does not proceed without a yes.
Two-question qualification, maximum. Current setup, then "is this a priority?" The answers are load-bearing: they feed the next step.
The imagine moment. Never explain the system. Make the prospect picture THEIR Monday morning with the problem gone, built from what they just said. The brain remembers images, not features.
Micro-pitch with one number, then stop talking. Silence is a tool.
The booking ask: always two slots, never one.
The two-path close (towing variant, restaurants): pair a low-friction yes (a recording, a demo SMS) with the high-value yes (the booked call). The easy path converts cold to warm; the hard path converts warm to pipeline.
The double-qualify rule (UK): a polite "sure, send it" is not interest. Get a genuine yes twice, and try to talk them out of it once in between. A clean no is a success; it stops you chasing ghosts.
The identity close, used once. "Picture two businesses six months from now..." Only on hesitation, never on a hard no.
Graceful exit. One soft pushback maximum, then mark not interested, suppress the number, and hang up politely. Restraint is the whole skill: the agent gatekeeps the meeting, it never sells the product, never negotiates price, and routes everything hard to the founder.
Honest AI disclosure whenever asked. Never claim to be human. It books better, and in some jurisdictions it is also the law. The restaurant pack goes further and discloses up front, because there the agent itself is the product demo.

Why the packs differ (per-market register tuning)

Dubai: formal-warm, quietly confident; excitement reads as desperation. Target 3-5 minutes, last-resort close at 6.
US brands: warm, never chummy; founders smell a hard sell in one sentence. Target 3-4 minutes, bail at 5. Faster market, earlier exit.
US towing: working-class peer voice, slower pace (the owner is under a truck), dollar math not vibe math, industry vocabulary mirrored, never recited.
Estonian restaurants: natural casual Estonian, AI disclosed in the opener as proof, 2-strike exit because the market is small and reputation compounds.
UK locals: understated and low-energy on purpose; the double-qualify gates filter the polite false yeses that UK courtesy produces.

What you need

An ElevenLabs account with Conversational AI enabled (separate from plain TTS).
A Twilio account and a number imported into ElevenLabs for outbound. Numbers local to the prospect picked up at roughly twice the rate in our US campaign, at about $1 per month per number.
For the towing and restaurant packs: a webhook backend (n8n, Make.com, or any HTTP service) for the booking, SMS, and suppression tools. The brand packs use the native Calendly integration; the UK pack needs no tools at all.
Cost ballpark from our US campaign: roughly $0.30 of ElevenLabs Conversational AI cost per 4-minute call on a Haiku-class LLM, plus about $0.04 of Twilio US outbound.

Before you dial: the legal part

Outbound AI calling is regulated, and in some places heavily. In the US, TCPA applies, and the FCC has ruled that AI-generated voices fall under its robocall rules: consent requirements, National Do-Not-Call scrubbing before every batch, calling hours of 8 AM to 9 PM in the prospect's local time, accurate caller ID, and recording disclosure in two-party-consent states. Elsewhere, check PECR in the UK and GDPR plus national telemarketing law in the EU. These packs default to warm-lead use: people who know you, opted in, or asked to hear from you. Cold use is on you. Check your jurisdiction and your lead list's consent status before the first dial, honor every opt-out immediately and permanently, and have your agent disclose honestly when asked. This is not legal advice.

Placeholder glossary

Placeholder	Meaning
`{{agent_name}}`	The name your agent introduces itself with. Name it in the first line.
`{{owner_name}}`	The human who takes the booked meetings.
`{{company_name}}`	Your company.
`{{case_study_client}}`	One real client proof story. Use a real one or cut the line.
`{{monthly_price}}`, `{{setup_price}}`, `{{setup_price_range}}`, `{{monthly_price_from}}`, `{{setup_price_floor}}`, `{{monthly_price_floor}}`, `{{mid_tier_monthly_range}}`	Your prices, written out in words (TTS reads digits awkwardly).
`{{your_booking_link}}`	Your Calendly / Cal.com event type.
`{{your_webhook_url}}`, `{{your_webhook_token}}`	Your webhook base URL and a random bearer token you generate.
`{{demo_base_url}}`, `{{demo_page_url_spoken}}`	Restaurant pack: your per-prospect demo page, and its spoken form for voicemail.
`{{sending_email_spoken}}`	UK pack: your sending address as it is spoken aloud.
`{{first_name}}`, `{{business_name}}`, `{{restaurant_name}}`, etc.	Per-lead dynamic variables your dialer fills at call time.

Built by kratt (krattconsultancy.com).

This is the exact system we run for clients. Take it. Or book the free audit and we run it for you. https://cal.eu/krattconsultancy/30min