Nothing in that video was recorded. Not the screens, not my voice, not me. The guy talking is an AI clone, and my camera stayed off the whole time.
I'm not showing this off. I'm showing you how it's built, because the pieces are cheaper and simpler than most people think. In this post: the four tools, how to clone your own voice and face, the one rule that stops you paying for the same render twice, the two ways to put your face on screen, and the full bill at the end.
The stack is four tools
Claude Code runs everything. It writes the script, calls the other tools, and stitches the final video together.
ElevenLabs handles the voice. You clone it once, and after that every line you type comes out sounding like you.
HeyGen makes the avatar. That's the digital twin.
And the graphics come from HyperFrames, which is the part almost nobody mentions. HyperFrames is HeyGen's own renderer, it turns an HTML page into video frame by frame, and it's open source under Apache 2.0. No watermarks, no render fees. They charge for the avatar and give the renderer away.
So out of the whole stack you pay for two things: the voice and the face. Everything else is free.
Cloning your voice
ElevenLabs needs a few minutes of clean audio. I gave it about two minutes of me talking into a mic, no music, no background noise, and that was enough. Instant voice cloning is on the Starter plan at $6 a month.
One thing I learned the hard way. After the first clone worked, I re-recorded my samples with a better mic in a quieter room, expecting a better clone. The new one lost. I ran both versions blind against four test lines and the old scrappy clone sounded more like me.
So don't get stuck chasing perfect samples. Clean and clear beats fancy.
Cloning your face
The face needs one thing from you: a single two-minute take, straight into the lens. No cuts, no reading off a script, just you talking naturally.
Three things make that take work:
- Keep your hands low and out of frame.
- Keep your head fairly still.
- Pause with your lips closed every few sentences.
That take is what HeyGen learns your face from, so the better it is, the better your twin. Then it asks you to record a consent video on your webcam where you read a short code out loud, which is how they stop people cloning a face that isn't theirs.
Training took a few minutes and creating the twin cost zero credits. The plan that makes it usable is Creator at $29 a month, and that includes the render credits.
The script is a list of beats, not a document
Once a copy of you exists that will read anything you type, the script stops being prose. It becomes a list of beats, and each beat says three things: who's on screen (avatar or graphic), the exact words spoken, and what the visual is doing while that happens.
Writing for a voice clone has its own rules:
- Spell out every number. The clone reads text, not math. Nineteen euros gets written as words, never as digits.
- Keep sentences short. Long ones drift out of rhythm.
- Use contractions, or it sounds like a robot reading a contract.
The pipeline lints all of this before anything gets generated, so you find the problems for free instead of finding them in a render you paid for.
Lock the script before you spend a cent
This is the rule that pays for itself.
Before any paid step runs, the script gets locked and the pipeline takes a fingerprint of it. From that moment, every paid stage refuses to run if a single word changed.
Why so strict? The avatar reads the script exactly as it was locked. Change a sentence after rendering and that render is garbage, so you're paying for it twice. Lock first, spend second.
Two ways to get your face on screen
Path one is the twin. Every segment of me in the video comes from that one two-minute training take. I upload audio from my voice clone and HeyGen makes the twin perform it. Worth knowing: HeyGen has its own built-in voice clone, and mine came back with an accent I don't have. Paid test, honest result. Skip their voice and bring your own audio.
Path two is a redub. You take real footage you already have, hand it new audio, and the AI resyncs your lips to match. No subscription, about $2.40 per minute. I paid for a real test, and it works: real footage of me saying words I never recorded.
My verdict after paying for both: up close, the redub just isn't as sharp as the twin. It gets soft when you blow it up, and soft is the first thing people notice. The twin costs a monthly plan and stays clean at full screen, so every segment of me in this video is the twin. The redub stays in my pocket for quick jobs on old footage.
The graphics build themselves
Every screen you see in the video was generated too. For each screen beat, Claude writes a small web page, and the animations are timed to my voice word by word using real timestamps from the audio rather than guesses. HyperFrames then renders that page into 4K video.
Claude drafts a few style options and I pick one. That pick is one of the few moments a human is genuinely needed.
Assembly, and the trick that sells it
Everything then goes through one assembly step: the voice, the avatars, the graphics. It carries audio across the cuts, upscales the avatar footage to 4K, and balances loudness to YouTube's standard.
The detail I like most is smaller than that. A cloned voice is generated in a perfect void, no room, no air. Real recordings always carry a room sound. So the pipeline lays a quiet room tone under the whole video. Without it your ear knows something is fake but can't tell you why.
The honest 20%
Anyone telling you that you press a button, walk away, and come back to a finished video is selling you something.
This pipeline automates about 80% of the work. The other 20% is still me. I approve the script, I pick the graphics style, and I watch the preview and catch what looks off. That 20% is the difference between a video and spam.
The pipeline does grade its own work first. It measures every clip against the plan, checks loudness, scans for dead air, and pulls frames from every avatar segment so I can eyeball the lip sync, then writes it all into a report. If a segment drifted, it fails before upload instead of after.
The bill
Everything you watched, with real numbers:
- HeyGen: $29 for the month, which covered every twin segment with credits to spare.
- ElevenLabs: $6 for the month, and the whole narration used about a quarter of the quota.
- The redub test: $1.65, paid per second.
- Graphics: $0.
- Assembly: $0.
Under $40 all in, and most of that is two subscriptions that also cover next month's videos.
Build it yourself
I packaged the whole thing as a Claude Code skill: the beats, the gates, the cost checks, all of it. It's €19 on Gumroad as the AI Video Pipeline kit, and you bring your own keys.
If you'd rather build it with me next to you, that's what the Build With AI Club is for. I'm in there every day answering questions, and you can try it free for 7 days.
One last thing worth saying out loud: if you publish a video with a synthetic version of a real person in it, turn on YouTube's altered content toggle at upload. I did. It's a disclosure, not a punishment.