I built a
podcast with an
AI voice in 24 hours.
Three narrative podcast episodes, "Business Wars" style. The voices are synthetic. The scripts were written by Claude sub-agents. The mixing is automated. Total cost: thirty-three dollars. Here's how, step by step, with nothing hidden from you.
What you'll discover
- Why I wanted to make a podcast without putting my own voice on the mic
- The topic I picked: the AI war (OpenAI, Anthropic, Musk) — three episodes in business-thriller mode
- How I picked the voice that dramatizes, listened to a five-minute test, and nearly broke everything with an invisible bug
- The system that turns a written script into a finished MP3: research → writing → synthetic voice → mixing
- The three episodes to listen to right below
- Transparency: article written with Claude (Anthropic), audio production orchestrated by Claude Code sub-agents + ElevenLabs, synthetic voices (public library, no voice cloned from a real person). If you spot a factual error, a misattributed voice, or an outdated number, write to me and I'll fix it.
I wanted to tell a story. Not mine.
On April 22, I'd started a personal podcast project that morning. Intimate format, three minutes, my own cloned voice and my wife Shirley's. I was going to tell real-life stories. The pilot was called "The Mac That Closes at Midnight."
That evening, I switched to something completely different. I want to make a narrative podcast Wondery style — you know, that American company that produces Business Wars or The Shrink Next Door. Fifteen-minute format, tense host voice, cinematic music, immersive sound effects. Except I don't want to put my voice on the mic or write it alone.
The idea: to test whether I can pull off a real pro podcast production while fully owning the AI — synthetic voices, scripts written by Claude sub-agents, automated mixing. Not on the sly. Not finishing it by hand. Truly end to end.
Why I'm telling this honestly
Everything that follows is true. The voices are synthetic (noted in the project's memory). The scripts were written by Claude sub-agents that I briefed and validated. The mixing runs in Python. My value-add was: choosing the topics, validating each step by ear, rejecting whatever didn't sound right, steering things back when the result drifted. The tone across the whole chain — that's me.
Three episodes on the AI war.
Wondery is, above all, about the human conflicts behind the big companies. Apple vs. Microsoft. Netflix vs. Blockbuster. So I look for a current tech conflict that deserves a thriller. Three candidates emerge:
- Sam Altman vs. Dario Amodei — the founding split. Dario leaves OpenAI in late 2020 with seven people, founds Anthropic, and five years later wages a commercial war on his former boss. Little known to the general public.
- The four days that nearly killed OpenAI — the coup of November 2023. Sam fired on a Friday, comes back in triumph on Tuesday. Pure drama.
- Elon Musk vs. Sam Altman — the ongoing lawsuit. Musk co-founded OpenAI, left furious in 2018, and has been suing OpenAI since 2024. He's claiming one hundred thirty-four billion dollars.
I decide to do all three. A trilogy that tells the AI war in chronological order. The three episodes talk to each other — a character who dies at the end of episode one comes back in glory in episode two, and takes the witness stand in episode three.
Listen to them right now.
Before we talk about how it's made, listen to the result. Each episode runs about fifteen minutes. The voices are synthetic but they don't sound robotic — they dramatize, they whisper, they cut each other off.
The Fracture
How Sam Altman's friend founded his rival. The night in December 2020 when seven people walked out of the most exciting company of the decade. Today their company is worth three hundred eighty billion.
The Four Days
How four people nearly made the most valuable company in tech implode, in four days. Friday, November 17, 2023, Sam Altman is fired by his own board. The following Tuesday, he comes back as boss.
Sworn Brothers
Two fathers tearing each other apart over custody of AGI. Elon Musk sues Sam Altman, claims one hundred thirty-four billion dollars, and jury selection begins in four real days as I publish this.
If you only listen to one
Start with episode two. It's the most dialogue-driven — a thriller across four days, seven voices answering each other. It's also the one where you can best hear what synthetic voices can do in dramatic mode.
Prefer to listen without reading?
Head to the podcast page — there you'll find all three episodes with a dedicated player, the covers, and everything you need to subscribe on Spotify, Apple Podcasts, or via RSS. This article here is the making-of. The podcast page is pure listening.
The system, in four steps.
For each episode, I repeated the same chain. I'll show you the main blocks. Under the hood it's more complex — but the idea boils down to four beats.
Research
I launch three or four Claude sub-agents in parallel on the topic. One goes after the factual history (sourced quotes, dates, scenes). Another looks at Wondery writing techniques. A third audits what works in synthetic voice. They write their results to files, I consolidate.
Writing
I brief a screenwriter sub-agent that reads the research results, follows a validated structure (cold open, three acts, climax, outro), and writes the script while respecting the synthetic-voice constraints (numbers spelled out, punctuation to drive the pauses, tone cues baked in).
Voice generation
A Python script walks through the script, figures out who's speaking, and calls the ElevenLabs API for each line. The narrator is generated in classic mode. The dialogue scenes (the trial, the negotiations) go through a special mode where several voices share the same take so the back-and-forth sounds natural.
Mixing
A Python assembler adds the background music, the sound effects at the right moments, the dramatic silences, and smooths out every transition. The whole thing then runs through a masterer that adjusts the volume to Spotify and Apple Podcasts standards. Output: an MP3 ready to publish.
For episode three, I added a fifth step: a quality control by two new sub-agents. One audits the sound effects and the music. The other reads the script like a first-time listener and flags anything that isn't clear. They saved me from a major bug — I'd used a character from episode two in the role of an anonymous journalist in episode three, which would have lost any loyal listener.
The five-minute test.
Before generating the full fifteen minutes of episode one, I wanted to validate the narrator's voice. It's the heart of a podcast — if the voice sounds bad, everything else is wrecked. I test two candidates I found in the ElevenLabs library: Theodore HQ and Paul K.
Theodore HQ is labeled "serene and grounded." Its profile says it's tuned for guided meditation. On paper that sounds warm. In reality, it's exactly the opposite of what I want. A meditation voice is trained to slow down your heart rate. Wondery wants to raise it. I reject Theodore.
Paul K is labeled "deep French narrator." Deep, warm voice, built for documentary narration. That's exactly the profile I'm after. I decide to test it on seventy-five seconds (the episode's cold open), for thirty cents of API cost.
First useful note: the pacing within a block is good, it's just the transitions that feel abrupt. We add seven-tenths of a second of silence between each block in the mix. Test redone. "And it's perfect now."
Five minutes of validation instead of a full regeneration at ten dollars. This is probably the most useful lesson of the whole project: test a small chunk before you fire up the machine.
The invisible bug.
I launch the full generation of episode one. Everything goes through. I do the Python mixing. I listen to the result. And then something bugs me. The music and the sound effects are perfect over the first fifteen seconds. After that… nothing. Just the voices.
The music disappeared after 20 seconds
The voices were recorded in mono (a single channel). The music in stereo (two channels). When the assembler mixed them, it aligned everything to the lowest common denominator — mono. The music, which lives in the left and right channels, got crushed the moment a voice took over.
Six lines of Python to force everything to stereo before mixing. The voices are up-converted (the same sound in both channels), the music stays intact, and everything mixes cleanly.
I would never have caught this bug without listening. The code passed. No error. Everything was technically correct. But to the ear, it was broken. This little episode changed how I work with AI from then on: a code audit alone is no substitute for a human audit through the senses.
Other, smaller bugs, the same day
The crossfade between two music tracks used the first few seconds of both tracks instead of the end of the old one and the start of the new one (a "cut" effect). The music bed restarted at sample zero on every voice line (a "stutter" effect). A sub-agent hallucinated a voice ID that didn't exist in the library. Every time, listening was the tiebreaker.
Three episodes, three leaps.
Each episode pushed me forward on one point. Here's what I added each time.
Episode 1 — Base pipeline
Narrator voice, three occasional characters, background music, sound effects. The classic Wondery grid. Discovery of the invisible bugs, mixing overhaul, validating the technical rules for what came next.
Episode 2 — More dialogue
You asked me to add more exchanges between the characters. Good idea. The episode went from six lines to thirty. I added four new characters (Ilya Sutskever, Helen Toner, Greg Brockman, Satya Nadella). It turned into a multi-camp thriller.
This is also when a sub-agent flagged a voice ID that didn't exist — a character I had to swap out in five minutes after a 404 error. Lesson learned: for the next episodes, validate any ID by generating a test before flagging it to me.
Episode 3 — Studio production
Three new things. The dialogue scenes switch to multi-voice mode (the API "hears" the previous turns and adapts the prosody, the interruptions are genuinely audible). The sound atmospheres become multi-layered — the courthouse lobby is three sounds mixed together (marble reverb, press murmurs, approaching footsteps). And each character gets their own sonic signature: Sam Altman announced by the clink of ice cubes, Sutskever by a leather chair creaking, Musk's lawyer by the snap of his binder.
I also set up an automated quality control: two sub-agents audit each episode before publishing. One checks that the sound effects match the script. The other reads it like a first-time listener and flags the fuzzy passages. It's exactly what Wondery editors do in the US. Except this is free.
What it really costs.
Total cost of all three episodes, from the first bit of research to the final master. Everything included: voice generation, sound effects, music, quality control. No extra gear. My laptop and an internet connection.
Wondery in the US spends several tens of thousands of dollars per episode — writing team, sound designer, voice actors, studio, music rights. I'm not saying my result matches theirs. I'm saying that for the price of a sandwich, I was able to produce something worth listening to. That's what changes.
And the time?
From the first idea to the master of the third episode: twenty-four real hours, on top of a normal day. Most of the time I was doing nothing — the sub-agents were running in the background. My real work was validating by ear, rejecting weak takes, and steering things back when the output drifted.
Five transferable lessons.
If you want to do the same — on a podcast, but also on any creative project where AI can do the heavy lifting — here's what I'd keep.
Testing a small chunk is worth its weight in gold. Five minutes of validation on seventy-five seconds saved me a full regeneration at ten dollars. Always run a mini-test before committing the machine to the big job.
The human ear sees what the code can't. The stereo bug would never have been caught by a code audit. On a creative project, the ear (or the eye, the palate, depending on the topic) stays the final judge.
Capitalizing on what you learn is the multiplier. I kept every rule validated in the first episode in memory. The second and third never rediscovered the same bugs. Without that, I'd have lost half the time over again each round.
Sub-agents in parallel are a game changer. Four independent research tasks in ten minutes instead of forty in series. The factor of four isn't a gimmick — it's what lets you do the trilogy in a day instead of a week.
Automated quality control takes you from good to pro. Having each episode read by two sub-agents playing critical editors is exactly what Wondery US does. Except here it's free. Don't pass it up.
Trust the human on creative calls, challenge them on poorly informed technical ones. When I wanted Theodore HQ, the agent checked, explained why it was a bad lead, and I changed my mind. When I wanted Paul K, straight validation. No yes-man, no knee-jerk contradiction.
Want to try it?
If you want to make your own podcast — or any creative project where AI can take on a big chunk of the work — here's what I'd suggest to start simple:
- Pick a topic you know. The quality of the result depends eighty percent on the quality of the material you feed the AI. If you don't know your topic, the AI will make things up.
- Start with a mini-test. A seventy-five-second cold open. Thirty cents. If it sounds bad, you've lost nothing. If it sounds good, you know you can fire up the machine.
- Capitalize on what works. Once a setting is validated by ear, write it down somewhere. The next episodes will go five times faster.
- Ask for a quality control before publishing. A sub-agent that reads it like a beginner is free, it's fast, and it keeps you from publishing something incomprehensible.
If you want the technical details (the Python code, the exact parameters, the prompts given to the sub-agents), sign up for AI Playbook — I'll walk you through, in a future edition, how I keep refining the system. If you just want to listen to the episodes properly (dedicated player, RSS / Spotify / Apple Podcasts subscription), the podcast page is made for that. And if you want to send me your take on the three episodes, I read everything. It's the human ear that decides.
To go further on the AI war itself, I'd point you to what the AI world looks like in 5, 10, 20 years — that's my forward-looking read on the players named in the trilogy. And if you want to see another making-of in the same spirit (a field sales tool built in two weeks with Claude Code), it's over here.
AI podcast FAQ.
How much does it really cost to produce an AI podcast episode?
For the pilot trilogy, the total cost was 33 dollars for all three episodes: $8 for ep 1, $10 for ep 2, $15 for ep 3. Everything included: ElevenLabs voice generation, music, sound effects, automated quality control. No extra gear. For a single 15-minute episode, count on $8 to $15 depending on how dialogue-heavy it is.
Which synthetic voice should you pick for dramatic narration?
On ElevenLabs, I tested Theodore HQ (labeled serene, tuned for meditation) and Paul K (deep French narrator). Theodore is trained to slow down your heart rate — exactly the opposite of a Wondery thriller. Paul K, a deep, warm voice built for documentary narration, got the green light. Golden rule: test 75 seconds for 30 cents before generating a full 15 minutes at $10.
How much actual human time does it take (vs. machine time)?
24 real hours between the first idea and the master of the third episode, on top of a normal day. Most of the time I was doing nothing — the sub-agents were running in the background. My real work (maybe 4 to 6 hours total) was validating by ear, rejecting weak takes, steering things back.
Is it legal to use synthetic voices to portray Sam Altman, Musk, etc.?
The voices used come from the public ElevenLabs library (Paul K, etc.) — no voice cloned from Sam Altman, Musk, or Sutskever. It's journalistic voice-over narration, not impersonation. The transparency is owned upfront, right from the hero of the making-of. The Scarlett Johansson vs. OpenAI Sky case in May 2024: absolutely avoid any voice that would resemble a real person without their consent.
What's the risk of AI hallucination in a narrative script?
Real. On the trilogy, a sub-agent hallucinated a voice ID that didn't exist (404 at generation time). In episode 3, I mistakenly used a character from episode 2 in the role of an anonymous journalist — a major bug caught by a quality-control sub-agent. Rule: have every script read by two critical sub-agents before publishing. For sensitive info (date, quote, number), check it yourself against a primary source.
Which tools should you use to produce an AI podcast in 2026?
The combo I tested: Claude Code to orchestrate the sub-agents, Anthropic Claude for the scripts, ElevenLabs for the voice (classic mode for narration, multi-voice mode for dialogue), Python for mixing and mastering to Spotify and Apple Podcasts standards, audio hosting on Cloudflare R2, distribution via an RSS feed.
How long will the MP3 files stay online?
The MP3s are hosted on Cloudflare R2 (free tier, 10 GB storage and 0 egress) for as long as the project lives. The RSS feed is submitted to Apple Podcasts and Spotify for Podcasters. As long as the Jerwis Productions channel stays active, the episodes keep being distributed. No expiration date planned.
Why not put your own real voice on the mic?
A deliberate editorial choice. The project wanted to test whether you can produce a Wondery-like format while fully owning the AI — synthetic voices, sub-agent scripts, automated mixing. Not on the sly, not finishing it by hand. Putting my voice in would have broken the demonstration. I have another pilot project (The Mac That Closes at Midnight) with my cloned voice and my wife's, but that's an intimate 3-minute format, not a 15-minute business thriller.
What's the exact Claude Code workflow?
Four phases plus one: research (3-4 sub-agents in parallel gathering history, Wondery writing techniques, a voice audit), writing (a screenwriter sub-agent that follows cold open / 3 acts / climax / outro), voice generation (Python calling ElevenLabs line by line), mixing (a Python assembler that adds music, sound effects, silences, transitions, then masters to Spotify-Apple standards). Phase 5 added in ep 3: quality control by 2 critical sub-agents (one on the sound effects, one on editorial clarity).
At what volume is it worth it vs. a voice actor?
Wondery US spends tens of thousands of dollars per episode (writing team, sound designer, voice actors, studio, music rights). A French voice actor for 15 minutes of narration runs between €300 and €800 before tax depending on their rate and the distribution. For a pilot or a side project, AI at $8-15 an episode is unbeatable. For a flagship monetized series, the voice actor still makes sense — AI is a complement, not a replacement.
Spot an error?
A misattributed voice, a wrong historical fact, a cost that's changed? Write to me at sagnier.jeremy@gmail.com · I fix it within 48h max and note the update date at the top. On sensitive topics (historical quotes, attribution of statements), your corrections are worth their weight in gold.
Want an episode 4?
I already have a few leads in mind: DeepSeek wiping out a trillion in market cap in a single day (January 2025) · Ilya Sutskever's departure from OpenAI (May 2024) · Google fumbling its own paper that invented language models · Scarlett Johansson vs. OpenAI over the Sky voice. If one of these topics grabs you — or if you have another one in mind — tell me.
Which topic do you want to hear?
Your vote helps me pick the next one. No sign-up needed, no email asked. Just your voice.
Shall we keep going?
I test AI for real and I share what works, no jargon, no hype. If this article helped you, the easiest way not to miss anything is my Friday letter. And if you have a question or a doubt: reply to me, I read everything.