10.01.26
The Core Idea
A voice agent builder turned his own phone chores into one typed sentence. A Claude Code skill reads the request, finds the businesses, pulls their numbers, and hands each call to a voice agent on Retell AI. The agent dials, works the phone menu, talks to a person, and either finishes the job or transfers the call to him. Claude reads the call logs afterward and reports what happened.
The demo
He typed one request: call three mechanics in Beckley, West Virginia, and get a quote for a brake pad replacement on a 2008 RAV4. The skill found three shops and placed three calls at once. Live call monitoring showed one call waiting on hold and two calls pressing keypad digits through automated phone menus.
On the call he played back, the shop asked for the VIN. The agent had none, asked for a ballpark price, heard no, and left a callback number in place of a quote. A human might have done no better, and the agent stayed polite and on task through an unplanned turn.
The honest scoreboard
Claude's report after the run: three calls placed, zero prices quoted, one shop promising a callback. The machine worked. The errand did not finish. Keep this gap in mind through every section below.
A DMV call. He sat on hold for more than 45 minutes, the line dropped, and he had to start over. He builds voice agents for clients for a living, so he built one for himself.
Claims made, not shown
Calls in multiple languages, hundreds of calls at the same time, negotiation on your behalf, and international dialing (his example: calling businesses in Japan). The video demonstrates three domestic calls and no negotiation.
One Agent. Two Jobs.
The voice agent does two things and nothing else. Every setting in the build serves one of the two.
Job one: finish the call
Transactional errands run start to finish without him: a quote, a status check, a booking, a cancellation. Backed by the press digit function, which taps keypad numbers to move through automated phone menus (IVR).
Job two: transfer to the owner
When the call needs a personal sign-off or details he prefers to give himself, the agent makes a warm transfer to his phone. Backed by the transfer call function. The prompt decides when, and only when needed.
The prompt, in five parts
Two Variables. One Skill.
Dynamic variables hold the information changing from call to call. A pizza order and an airline call need different requests, so the prompt stays fixed and two slots change.
caller_name
Who you are trying to reach. Tell Claude to call the DMV, and Claude fills the slot with the DMV on its own.
caller_purpose
Claude condenses your full request into a short brief. The prompt tells the agent to treat the brief as a guide, not a script, and break the brief into natural spoken chunks so the call does not sound read aloud.
The skill
The Claude Code skill comes from a shared Google Doc. You paste the skill into Claude Code with your API key. The skill takes a plain language request (his examples: order a large pepperoni from Domino's, ask Qatar Airways about a flight status), pulls out three values (who, why, and the number to dial), finds the number, and places the outbound call through the Retell API. Once the call ends, the skill uses the same key to read the call log and report results.
No webhook, on purpose
He skips the agent level webhook URL. Claude never listens to live call events. Claude checks the call logs after the fact, which keeps the whole build simple.
Settings And Setup
Settings, as he runs them
| Setting | His choice |
|---|---|
| Model | GPT 5.6, picked after his own testing |
| Voice | Grace, expressive mode V3, chosen for realism |
| Functions | Transfer call (warm transfer to him) and press digit (IVR navigation) |
| Knowledge base | None |
| Speech settings | Defaults |
| Webhook | None. Everything runs through the API. |
Setup in six steps
The transcript spells the platform "Retail AI." The product is Retell AI.
Hand Off
The Call.
Core message
You lose an hour a week to calls a script would run, and you put off the rest. Sort your calls into errands and decisions. Give the errands to a voice agent with a written brief. Keep every decision for yourself. Your calls get made the same day, and your phone rings only when the call needs you.
Target viewer
Early career, solo or junior. No assistant, no office manager, a phone list growing faster than the free time to work through the list.
The Calls Stack Up
You started the job, the apartment, the business. Now the calls stack up. Three quotes for the office move. A billing error on the internet account. A dentist reschedule. A vendor who never confirmed Thursday.
Each call takes ten minutes of work and thirty minutes of hold music, so you push each one to tomorrow. Tomorrow fills too. The standard advice says block an hour and power through. You try, and one transfer loop eats the hour.
The problem was never discipline. You treat every call as yours, when most calls are errands with a known goal and a factual answer. Errands follow a script. Scripts are delegable.
The source creator hit the same wall at the DMV and built a caller from parts you rent by the minute: a voice agent, a phone number, and one Claude Code skill. His own demo also shows the catch. Three calls went out and zero prices came back, because a shop asked a question the agent had no answer for. The build takes an afternoon. The brief is the job.
Count last week's calls. How many needed your judgment, and how many needed only your patience?
The five moves ahead
Sort The Calls Before You Build The Caller
An agent belongs on some calls and nowhere near others. Log one week of calls you made or avoided. Mark each one E for errand or D for decision.
An errand has a fixed goal and a factual answer: a price, a date, a status, a booking, a cancellation. A decision carries stakes you set in the moment: a negotiation over your money, a call to a client, anything needing your account details or your signature.
The source agent runs on the same split. Transactional calls run end to end. Anything needing a sign-off or personal information goes to the owner. Your log tells you which pile holds most of your week.
Your week, sorted (sample)
| Call | Pile | Why |
|---|---|---|
| Get three movers' quotes | E | Fixed question, factual answer |
| Reschedule a dental cleaning | E | Known goal, two acceptable dates |
| Confirm a vendor delivery window | E | Status check |
| Dispute a $140 charge | D | Money plus account details |
| Ask the landlord about renewal terms | D | Negotiation with long term stakes |
| Call a prospect back | D | Relationship, your voice |
Start with one errand type you repeat at least twice a month. Quotes and appointment changes win most often, because the script barely changes between calls.
Write The Brief Before You Dial
The demo stalled on one question: what is the VIN? The agent had no answer, so the call ended with a callback in place of a price. Your agent meets the same wall on every call where you skip prep.
The source skill sends two variables: who to call and why. Those two get the agent through the phone menu. They do not get the errand finished. Add five fields to every brief.
What are the first two questions this business will ask? Does my HAVE READY line answer both? If no, the call is not ready.
Give The Agent A Job, A Limit, And Examples
The source prompt carries three load-bearing parts: a role, a primary objective, and example calls. Keep all three. Add a limit.
Role and disclosure
One line naming who the agent is and whom the agent works for. Add a disclosure rule: the agent identifies as an AI assistant when asked, and up front wherever your state requires.
Objective
Finish the errand on this call and reach a person. The source creator credits this line with making the agent push through menus and hold queues.
Limit
The source objective tells the agent to stay on until the task is resolved or every option is gone. Without a cap, the agent burns paid minutes in a queue you would have left. Write the cap in numbers.
End the call and log the result after any one of these: 15 minutes on hold. Two failed transfers inside the business. One callback promise with a named person and a time.
Examples
Write three short sample calls from your own errand list. One clean win. One where the business needs information you lack, showing the agent asking for a ballpark and then a callback. One where the call transfers to you. Match the examples to the errand type you picked in Point 01.
Hand Off The Errand. Keep The Decision.
The transfer function is the most important line in the build. The source agent transfers when a call needs a personal sign-off or details the owner prefers to give himself. Turn the feeling into a written list.
Transfer triggers
- The business asks for payment, a deposit, or a card number.
- A price lands above your ceiling. Write the number in the brief.
- The business asks for you by name.
- The call needs your signature, login, or an ID number.
- The other side offers a term not in your brief: a contract length, a cancellation fee, a package upgrade.
Make the handoff warm
A warm transfer means the agent briefs you before connecting the caller, so you pick up knowing the business, the quote so far, and the open question. Save the agent's number in your contacts so the transfer rings through focus modes and call screening.
Set the agent's authority
Decide what the agent agrees to without you. A reasonable starting point: confirm a date inside your listed options, accept a quote as information, and request a written estimate by email. Everything else waits for you.
Picture a new intern on the call. If you would not let the intern agree to the term, the agent transfers.
Read The Log, Not The Summary
Claude's report after the demo was honest, and still a summary. Summaries flatten the detail you need, like which shop asked for the VIN and who promised to call back.
Open the transcripts worth opening
Read the full transcript of any call ending without the DONE WHEN result. The miss is almost always in the brief: a missing fact, a vague purpose, a trigger the agent misread.
Track every call in one sheet
| Business | Outcome | Number quoted | Next action | Date |
|---|---|---|---|---|
| One row per call | Quote, callback, no answer, transferred | Price or range, blank if none | Book, retry, wait, drop | Call date |
The agent hands out your callback number, so callbacks ring you, not the agent. Log each callback in the same row the moment the call ends.
Measure before you scale
Run five calls of one errand type. Write down your manual minutes per call, hold time included. Then time your own work per agent call: writing the brief plus reading the log. Keep the agent where your review minutes beat your dial minutes.
The errand ladder
Once one errand type passes, climb one rung at a time: quotes, then appointment changes, then status checks, then cancellations with a refund. Each rung gets its own brief template and three example calls.
Three Quotes By Friday
A junior office coordinator needs three quotes for after-hours carpet cleaning in a 2,000 square foot office before Friday. Here is the full brief, typed into Claude once.
Call three commercial carpet cleaners near our office and get quotes. CALLER NAME: [business name] PURPOSE: Price to clean 2,000 sq ft of low-pile office carpet on a weekday after 6 p.m., plus the earliest available date. HAVE READY: 2,000 sq ft. Low-pile commercial carpet. One open floor, two small conference rooms. Weekdays after 6 p.m. Parking in the rear lot. DONE WHEN: A price or price range plus the earliest date, or a named person and a callback time. TRANSFER IF: They ask for a deposit, a signed agreement, or a card number, or the price tops $600. NEVER GIVE: Payment details or company account numbers. STOP AFTER: 15 minutes on hold, two failed transfers, or one named callback promise. Log the result and end the call.
The sheet after the run (sample values)
| Business | Outcome | Quoted | Next action | Date |
|---|---|---|---|---|
| Cleaner A | Quote | $340 | Book after manager OK | 10/01 |
| Cleaner B | Callback | None yet | Expect Dana by Thursday noon | 10/01 |
| Cleaner C | No answer | None | Retry Thursday 9 a.m. ET | 10/01 |
The time math (planning estimate)
By hand: three calls at about 20 minutes each with hold time, roughly 60 minutes. With the agent: about 3 minutes to adapt the brief and 5 minutes to read the logs, roughly 8 minutes of your attention. Replace both numbers with your own after the first five runs.
What The Video Skips
The build works. Eight gaps sit between a working build and errands you trust to run without you.
One Errand.
Seven Days.
Pick the call you keep pushing to tomorrow. Give the agent the errand. Keep every decision. By day seven you know your real minutes saved, and your brief gets better with every transcript you read.
Keep the decision.
10.01.26
Join The AigentLab
Every field guide starts a conversation. Inside the AigentLab you build alongside other solo operators, share your briefs and transcripts, and get feedback on the systems you run.
Bring your first errand. Leave with an agent running the call.