Answer every call.Book the right ones.
50 voice agents delivered as complete systems — system prompt, conversation flow, qualification logic, objection playbook, handoff rules and the 9 ways a live call actually goes wrong.
01 / The problem
A voice agent does not fail the way a document fails.
A bad email is deleted. A bad call is answered by a person who then remembers it — and, in three of the four regimes below, is recorded while it happens. These are the failures every card in this product is built around. They are not opinions about the market; they are what happens on a live line.
None of that is fixed by a better persona prompt. It is fixed by holding a slot before you speak it, scrubbing suppression at dial time rather than overnight, putting the disclosure before the first question, and testing all 9 of them before an agent gets a phone number. That is what the 50 cards do.
02 / The spine
Every agent has all six stages. No exceptions.
Most AI voice content stops at the persona. That is stage three of six. The stages either side of it — what makes the call happen, what the agent is allowed to know, what it writes back, and how you tell whether it worked — are where the result actually comes from, and they are the parts nobody writes down.
03 / The mechanic
Fill in one profile. All 61 prompts already know your business.
This is the part a prompt pack structurally cannot do, and the reason the offline app is in the box rather than a folder of text files.
- You fill in 7 fields
What you sell, your services and prices, your hours, your markets and your proof. Ten minutes, once — and if you own the other two Vaults, 12 more import themselves, about 55% of what is required.
- Agents write twelve more
Your qualification questions and rubric, the objection playbook, the knowledge base and refusal list, the booking rules, the transfer map, the CRM field map and the pronunciation lexicon get written by the agents as you run them. The profile fills itself in.
- Five are guardrails, shipped pre-filled
Disclosure line, consent basis, suppression source, recording policy and banned phrases. They need your review, not your invention — and VA-44 is the card that reviews them with you.
- Every prompt carries it automatically
59 typed variables —
{{VOICE_CORE}},{{DISCLOSURE}},{{NEVER_CLAIM}}. Hit copy and your context is already inside, so agent 37 already knows what agents 6, 8 and 21 established.
What a prompt pack hands you
You are a friendly receptionist for
a business. Book appointments.What one click copies here
[your business, what you sell, your
services and prices, your hours and
timezone, your markets, the agent's
persona and tone card, the qualification
rubric, the objection playbook, the
knowledge base, the refusal list, the
booking rules, the handoff rules, the CRM
field map, the disclosure line, the never-
claim list and the banned phrases - all
already filled in from your Voice Core]
You are calling a lead who asked us for
something and did not book. Say the
disclosure first, before any question.04 / What it actually does
One cancellation, four minutes, and a slot that never went empty.
This is the worked example from VA-46, printed in full in the book. The practice is fictional and the calls are simulated — but the arithmetic is real, and an independent test recomputes it on every build.
then release the old,
then refill →
Nobody worked harder. The same cancellation, run through a chain that knows the order of operations, recovered £150 of capacity for 29p of call time — and note the order at 14:03: new slot committed, then the old one released, then the refill fired. Reverse the first two and the patient briefly has no appointment. Skip the third and Friday is empty anyway.
That is one agent of 50. Every one carries a worked example like this, with the transcript shown, the arithmetic recomputed and the failure modes named.
Get the Vault →05 / Anatomy
Thirteen blocks on every card. A card missing one does not ship.
Enforced by a linter, not by eye — all 50 cards are checked automatically before the book is built, including four sub-blocks the linter treats as non-negotiable: qualification logic that separates hard gates from weighted scores, the objection playbook, the handoff rules and the CRM write schema. Averaging 4,073 words each.
One sentence, in your words — what this agent is on the phone to do.
A call, an event or a threshold. Never “whenever.”
Each input: what it is, where it comes from, and the CRM write schema on the way out.
Platform-agnostic. Vapi, Retell, Bland, ElevenLabs, Synthflow or your own — named, not assumed.
This agent’s six stages drawn with its real branches and loops.
System prompt, conversation flow, qualification logic, objection playbook and handoff rules.
A simulated call, labelled as one, with the transcript, the log and the arithmetic.
Persona, pace, attempt ceiling, quiet hours, market — and the exact line to edit.
What to reject, as checkboxes, plus the compliance surface for this agent.
One KPI, its formula, how to baseline it, and what it cannot prove.
Minutes by hand versus by agent, and the cost per call at published rates.
All 9 voice failure modes answered — barge-in, dead air, mishear, voicemail, IVR, hallucinated slots, the robot question, latency, consent withdrawn.
Manual → semi → full, with the system that automates it named.
06 / The test lab
An agent that has not passed does not get a phone number.
41 test calls across seven groups — opening, disclosure, comprehension, qualification, objections, booking and escalation — scored on 7 axes. Compliance is not averaged with the rest: anything below full marks is a fail regardless of the mean, because a disclosure that works four times in five is a disclosure that does not work. VA-48 re-runs the whole set weekly, because agents drift every time somebody edits a prompt to make it sound warmer.
The tab you open in the morning
Every agent you have deployed, what it is doing, its connect, qualification, booking and escalation rates, its cost per booking, and when it was last tested. It tells you when an agent has not been tested in a fortnight.
Eleven questions to a spec
Answer eleven questions and it assembles the system prompt, the conversation flow, the qualification logic and the config blocks — with your Voice Core already inside.
Drift is the real failure
The example in the book is an edit that made an agent measurably warmer and silently broke its hold-release step. Naturalness went up one point; booking went down two. Nothing else would have noticed.
07 / The book
740 pages, and every agent is drawn.
Each agent gets its own spine diagram with its real branches and loops — 50 of them, drawn as vectors so they stay sharp when you print. Prompts are set in a typeface chosen so that copy-and-paste out of the PDF gives you the prompt back exactly, and a build test checks every placeholder survives into the text layer.


08 / Contents
All 50, named before you buy.
Not “50+ agents” with a vague list. Here is every one, and what it produces.
Inbound & Lead Response 10
- 01Instant Website Lead Callback AgentCalled, qualified lead
- 02Inbound Lead Qualification AgentQualification verdict
- 03High-Intent Lead Callback AgentPriority callback
- 04Demo Request Qualification AgentQualified demo booking
- 05Consultation Booking AgentBooked consultation
- 06Multi-Question Deep Qualification AgentFull qualification record
- 07Budget & Timeline Qualification AgentBudget and timing record
- 08Lead Score & Human Handoff AgentScored lead, routed
- 09Paid-Ad Speed-to-Lead AgentSub-60s first contact
- 10Inbound Call Router & Intent ClassifierClassified, routed call
Appointment & Scheduling 10
- 11AI Appointment Booking AgentConfirmed appointment
- 12AI Appointment Reminder AgentConfirmed or rescheduled
- 13AI No-Show Recovery AgentRecovered booking
- 14AI Rescheduling AgentMoved appointment
- 15AI Demo Booking AgentBooked demo
- 16AI Property Viewing Booking AgentBooked viewing
- 17AI Consultation Follow-Up AgentPost-consult next step
- 18Waitlist & Cancellation Fill AgentRefilled slot
- 19Availability Truth & Slot-Hold AgentVerified, held slot
- 20Multi-Location & Provider Routing AgentCorrectly routed booking
Sales & Follow-Up 10
- 21Cold Lead Re-Engagement AgentRevived lead
- 22Warm Lead Follow-Up AgentAdvanced lead
- 23Post-Demo Follow-Up AgentPost-demo decision
- 24Proposal Follow-Up AgentProposal decision
- 25Stalled Deal Recovery AgentRecovery outcome
- 26Lost Lead Win-Back AgentWin-back outcome
- 27Existing Customer Upsell AgentUpsell conversation
- 28Renewal & Retention AgentRenewal commitment
- 29Payment Reminder & Collections AgentPayment commitment
- 30Quote & Estimate Confirmation AgentAccepted or revised quote
Marketing, Service & Reputation 10
- 31Lead Magnet Follow-Up AgentQualified downloader
- 32Webinar Registration Confirmation AgentConfirmed attendance
- 33Event Attendance Confirmation AgentConfirmed attendance
- 34Customer Feedback Collection AgentStructured feedback
- 35Review Generation & Reputation AgentReview or private escalation
- 36Order Status & Delivery Update AgentResolved status query
- 37FAQ & Support Deflection AgentAnswered or escalated
- 38Outbound NPS & CSAT Survey AgentScored survey response
- 39Multilingual & Code-Switch AgentCall handled in-language
- 40Recruitment & Candidate Screening AgentScreened candidate
Advanced Voice Systems 10
- 41Missed Call Recovery SystemRecovered missed call
- 42AI Sales ReceptionistHandled inbound call
- 43Full AI Lead Conversion SystemAd click to booked meeting
- 44Voice Compliance & Consent GateDial/fix/block verdict
- 45Call to Transcript to CRM to Follow-UpLogged call, drafted next
- 46Booking to Reminder to No-Show to RefillProtected calendar
- 47Review & Reputation EngineReviews and escalations
- 48Voice Agent QA & Regression HarnessDrift report per agent
- 49Human Handoff & Live Transfer SystemWarm transfer or callback
- 50Voice Revenue Intelligence SystemVoice revenue intelligence
09 / Where to start
You never face a wall of 50 things.
Pick the track that sounds like you and you get eight agents in build order. Every track starts with the same card — VA-44, the compliance gate — because what it makes you add changes the first ten seconds of every conversation, and running it afterwards means re-tuning a flow you already spent a week on.
Local service business
Stop losing the phone: missed calls, bookings, reminders, reviews
Agency / voice-agent builder
Spec, build, test, deploy and report - for clients, on a retainer
Founder-led sales
One receptionist and one qualifier, without hiring anybody
Sales team / SDR
Speed to lead, qualification, booking, clean human handoff
Ops / CX manager
Routing, escalation, QA, regression and unit economics
10 / Automation
Deep on n8n. Honest about the rest.
Automation is the most attractive part of a product like this and the easiest place to overpromise. Here is exactly what you get.
n8n — 10 working systems
Real JSON files you import and run. Not screenshots of a canvas.
- Each one was imported into a live n8n instance and confirmed to round-trip before shipping
- Every system that can dial consults the VA-44 compliance gate first and fails closed
- Suppression is scrubbed at dial time; every calendar hold is released or expires
- 4 of the ten stop at a human approval gate before anything reaches a customer
- Every credential slot documented; no keys baked into the files
Make and Zapier — written setups
Step-by-step configuration with field maps, not import files.
- Module-by-module instructions with the exact fields to map
- Why: an import file for these platforms breaks the moment either changes a connector, and a stale file is worse than none
11 / The download
Everything in the zip, listed before you buy.
26 workbooks that calculate
The testing scorecard, the compliance and consent register, the DNC control sheet, the call-cost unit economics, the no-show refill log. Every formula was recomputed independently in Python and checked against the worked examples in the book.
The whole thing as a workspace
53 pages and 20 databases that import into Notion in one go, if that is where you'd rather live. The spreadsheets import into Sheets too.
Nothing phones home
The Command Center is a single HTML file. No account, no login, no server, no network calls. Your Voice Core stays in your browser. Open it from a USB stick in ten years and it still works.
12 / Straight talk
What this is not.
Worth being explicit, so you buy the right thing:
- Not a voice platform. It does not place calls, hold a phone number or transcribe anything. It is the specification you take to Vapi, Retell, Bland, ElevenLabs, Synthflow or your own stack — deliberately, so you are not locked to whichever vendor was cheapest the month this was written.
- Not legal advice. The compliance material names the regimes your answer depends on and gives you the gate that walks you through them. Confirming your position in each market you dial is yours to do, with a qualified adviser.
- Not a done-for-you service. These are agents you build. The first pass at each one takes real time — most cards state the manual baseline in minutes so you can see it before you start.
- Not a guarantee of results. There are no earnings claims here and no case studies, because I am not going to invent either. Every transcript in the book is a simulated call and says so. What's guaranteed is the method, the measurement, and that anything actually broken gets fixed.
- Not fully automated out of the box. 10 of the 50 have importable n8n systems. The rest are built and run by you, and each card tells you honestly which it is.
- Third of three, and it stands alone. It pairs with the Marketing and Sales Vaults — same architecture, same fictional businesses, and it imports both their profiles — but it needs neither. If you only ever buy this one, nothing here is missing.
13 / Licence
You buy it once. You own the files.
What you may do
- Use everything for your own business
- Use it on unlimited client projects, including paid work
- Adapt, edit and extend anything in it
- Keep it forever — no expiry, no seat check
- Take every future update, free
What you may not do
- Resell or redistribute the files
- Publish the contents in a course or members' area
- Repackage it as your own product
- Share one licence across a team — one person, one licence
Sales are final — and here is the trade. This is intellectual property delivered as a download. The moment the files are on your machine you hold the entire product, so it cannot be returned and there are no refunds. In exchange you get me: if anything is broken — a formula that miscalculates, a blueprint that will not import, a file that will not open — email prashant@thecodesmith.tech and I will fix it and reissue at no charge, along with free updates for life. So that the decision happens before you pay rather than after, all 50 agent names, the full file tree, the 13 blocks and a whole “what this is not” section are published above.
14 / The set
It already knows what the other two told it.
Three Vaults, one architecture. The Marketing Vault runs the half of the funnel that happens before a lead has a name; the Sales Vault runs the selling; this one runs the conversations. It is the only one of the three that reads from both — 12 Voice Core fields arrive already filled, about 55% of what is required.
AI Marketing Automation Vault
50 marketing workflows: positioning, content, campaigns, funnels, reporting. Fills the Brand Core.
AI Sales Automation Vault
50 sales workflows: prospecting, outreach, conversations, pipeline, forecast. Fills the Sales Core.
AI Voice Agent Automation Vault
50 voice agents: inbound, booking, follow-up, service and reputation. Reads both the others into its Voice Core.
Bought separately they are three complete systems. Bought together they are one business described once — and the handoff that usually loses the plot, between what marketing says, what sales sells and what the person on the phone actually promises, is the part that stops being manual.
15 / Get it
The whole Vault, one payment.
one-time purchase · no subscription
16 / Questions
Answers before you ask.
Is this just voice-agent prompts with better packaging?
No, and the difference is structural. A prompt pack gives you a persona to paste. Every agent here carries a trigger, a data contract, a conversation flow, qualification logic that separates hard gates from weighted scores, an objection playbook, handoff rules, a CRM write schema, a QA checklist, one KPI with its formula, and all 9 ways a live call goes wrong. That is why 203,000 words sit behind 50 agents — the instructions are the product. The prompt is one of thirteen blocks.
Which voice platform does this lock me into?
None. Every prompt is plain text and every card names a free path and a paid path — Vapi, Retell, Bland, ElevenLabs, Synthflow, or whatever you already run. The one thing the cards do assume is that your platform can detect an answering machine, cut audio on barge-in, and give you a transcript. Where a platform cannot, the card says what breaks.
What is the Voice Core, really?
One profile you fill in once: what you sell, your services and prices, your hours, your markets, your proof, and the things the agent must never say. 24 fields, but only 7 are yours to type — twelve get written by agents as you run them, and the rest are guardrails that ship pre-filled for your review. All 61 prompts are written against it using 59 typed variables, and the Command Center injects it every time you copy one.
I own the Marketing and Sales Vaults. Does this reuse them?
Yes, and it is the only one of the three that reads from both. 12 Voice Core fields import from the Marketing Vault’s Brand Core and the Sales Vault’s Sales Core — about 55% of what is required, filled before you type anything. Where the two overlap the Sales Vault wins, because its wording was written for a conversation. Neither of those products changes.
Is any of this legal advice?
No, and the product says so in several places. Voice agents sit on top of telecoms, consent, recording and AI-transparency law that differs by country and, in the United States, by state and by federal circuit. What the Vault does is name the regimes your answer depends on — GDPR, the TCPA and the FCC’s AI-voice ruling, Article 50 of the EU AI Act, India’s DLT rules — and give you VA-44, the gate that walks you through settling it for your markets before anything rings a phone. The confirmation itself belongs with a qualified adviser.
Are the transcripts in the worked examples real calls?
They are simulated, labelled as simulated everywhere they appear, and generated by running the shipped prompts against a model playing the caller. The businesses are fictional, and the phone numbers come from Ofcom’s ranges reserved for drama so nothing in the book can dial a real person. The costs are not estimates: they are arithmetic from published per-minute rates, and a test recomputes every one of them on each build.
How do I know an agent is good before it takes a real call?
That is what the Test Lab is for. 41 test calls across seven groups — opening, disclosure, comprehension, qualification, objections, booking and escalation — scored on 7 axes, with compliance treated as a pass/fail rather than an average. VA-48 then re-runs the whole set weekly, because an agent degrades silently every time somebody edits a prompt to make it sound warmer.
Will the n8n systems run without me?
The 10 files import and run on a schedule, yes — each was imported into a real n8n instance and confirmed to round-trip before shipping. But nothing in Section E dials a stranger unattended: every system that can place a call consults the compliance gate first and fails closed, and 4 of the ten stop at a human approval gate. Run the manual version first, and point them at a test number before a customer.
Can I use this for client work?
Yes — unlimited client projects, including paid work. One of the five tracks is written for exactly that: spec, build, test, deploy and report on a retainer. What you cannot do is resell or republish the files themselves. The licence is one page and says exactly this.
Can I get a refund?
No. The files are the product, and once they are downloaded they cannot be handed back, so sales are final. That is exactly why the whole contents are published on this page — all 50 agent names, the file tree, the 13 blocks and a full “what this is not” section — so the decision happens before you pay, not after.
What if something is broken?
Email me and I will fix it. A formula that miscalculates, a blueprint that will not import, a file that will not open, a prompt that has gone stale against a model update — that is on me, and the fix is reissued to everyone at no charge. Free updates for life are part of the same promise.
Every missed call was somebody who wanted something.
They rang once. Nobody counted it, because nothing visibly went wrong — no bounce, no error, no row in a report. The AI Voice Agent Automation Vault is $49.99, once, and it is downloadable in the next sixty seconds.