He Rambled Into His Phone

Talking points + backend specs

Part one is what each beat of the video covers, so you know the concepts without opening the deck. Part two is how it works on the backend in plain language — what a viewer has to go sign up for versus what Claude Code handles for them.

Part one — what each beat covers

Nine beats, in order. These are the concepts on screen; say them however you want. The tenth — the only CTA — is at the bottom of this page.

Beat 1 · cold open

Raw input vs. finished output

2 minutes 20 seconds of unscripted phone audio on one side
1 minute 43 seconds of finished, branded, narrated video on the other
No script, no prep — he hit record and talked about compliance

Beat 2 · the promise

No camera. No editor. No video software.

Built from a voice memo and a folder of text files
You're going to see every piece of how it works
Nobody is writing code by hand; you describe what you want in English
No mention of the free download here — the only CTA in the video is the last thing you say

Beat 3 · the raw material

It didn't rewrite him — it cleaned him up

Grammar tightened, the doubled-back sentence untangled, the sign-off cut
Every warning stayed exactly as strong as he made it — nothing softened
House terms go in as a standing rule the machine enforces on every video, automatically
If it had rewritten him it would sound like a robot wrote it; it still sounds like him

Beat 4 · the brand kit

Why it looks like your company, not a template

One stylesheet: colors, fonts, the shape of a card, the glow on a badge
Real assets — the actual logo, real screenshots of the actual app
Written once. Every video after that inherits it for free.
Honest: this is the part that took the most back-and-forth. Not one prompt.

Beat 5 · the voice

Two script files, one word different

One copy for humans (captions come from it), one copy only the voice engine reads
Text-to-speech says "lead" like the metal, every time — so the robot's copy is respelled leed
Captions on screen still show the correct spelling
Same voice across all three videos so they sound like one product

Beat 6 · the timing — this is the twist

It listens to its own voice

Normally animation means a timeline you drag and nudge — that's what eats the day
Instead: the generated voiceover gets transcribed back to get the exact millisecond of every word
Slides are pinned to words, not a clock — a slide starts when a phrase is spoken
Change the script, re-run it, and the whole video re-times itself
Honest gotcha: the transcriber writes numbers as digits, so a slide waiting on spelled-out numbers never fires

Beat 7 · the render

It can't drift out of sync

The video is a web page; a browser with no window jumps to a timestamp and screenshots it, thousands of times
Frames get glued to the voice, a music bed sits underneath, whooshes on the cuts
It isn't recording a screen — it's computing what the screen looks like at each moment
Slow machine, fast machine, doesn't matter: output time always equals source time

Beat 8 · the payoff

The thing you build is the next video

The first one was hours of real work
The second and third were new scripts into the same machine
Three finished videos in one working session, most of it spent on video one
Changes don't mean re-recording or re-editing — you change a text file and run it again

Beat 9 · where it doesn't work

This doesn't replace you

Not a talking head — if the job is trust, put a person on camera
Not a software demo — sometimes people need to watch the real clicks
The first build is genuinely work; leverage shows up on video two
It won't fix a bad explanation — you get a beautifully animated mess
What it's great at: the explainer in the middle nobody ever gets around to making

Nine beats, and none of them mention the download. The CTA is the tenth and final thing in the video — it's at the bottom of this page, where it belongs.

Part two — how it's built, in normal language

The only thing that matters for the video: what a viewer has to go sign up for, versus what Claude Code just handles for them.

The two things they have to go get

These are outside services with their own accounts and their own bills. Worth explaining on camera, because if someone skips these, nothing works.

Sign-up required · the one that matters most

ElevenLabs — this is the narrator

It's the AI voice. You pick a voice out of their library, and it reads your script out loud.
Paid account. You're billed on how much speech you generate, so a few short videos is not a big spend — but it is a real subscription, not free.
You copy an API key out of their dashboard once and hand it to Claude Code. That's the entire setup.
Pick one voice and never change it. Same voice on every video is what makes a set of videos feel like one product instead of three unrelated things.
The catch worth saying out loud: it mispronounces industry words. Ours reads "lead" like the metal, every single time. That's why there are two script files — one spelled correctly for the captions, one spelled phonetically just for the voice.

Sign-up required · cheap but essential

OpenAI — this is the timing

Uses their transcription model, Whisper. It listens to the AI voiceover and reports back the exact moment every single word is spoken.
Also just an API key you paste in once. Billed by how much audio you transcribe, and it's the cheaper of the two by a wide margin.
This is the least impressive-sounding piece and it's the reason the whole thing works — without word-level timing you're back to dragging things around a timeline.
Important if someone tries this: it has to give timings per word. Plenty of transcription tools only give you timings per sentence, and that's useless here.

What Claude Code handles — they never touch this

Mention these in passing so people know they exist, then move on. No accounts, no keys, no signup, no cost. Claude installs and drives all of it.

Thing	What it is	Do they do anything?
Playwright	Free open-source software that runs an invisible browser. It's what takes the thousands of screenshots.	No. It's not part of Claude Code — Claude installs it and runs it for you.
ffmpeg	Free, ancient, universal video tool. Glues the screenshots and the audio into an MP4.	No. Same deal.
The engine	The code that decides when each slide appears based on the voice.	No. Claude writes it.
The design system	One file holding your colors, fonts, and card styles.	You give Claude your colors, fonts, and logo once. It writes the file.

The line that keeps people watching: you're not writing any of this. You're describing what you want in plain English and looking at what comes back. The only genuinely manual steps are making two accounts and pasting two keys.

What it's actually doing, start to finish

1. Cleans up the rambling transcript into a script 2. ElevenLabs reads that script out loud ← their account 3. Whisper listens back and time-stamps every word ← their account 4. Builds the slides as a web page, each one pinned to a spoken phrase 5. An invisible browser photographs that page, frame by frame 6. ffmpeg glues the frames to the voice, drops music underneath, exports the MP4

The three ideas worth actually explaining

Idea	Why it matters
Two script files	One for humans, one spelled phonetically for the voice. It's the difference between "pretty good" and something you'd actually publish.
Pinned to words, not a clock	Slides are anchored to spoken phrases. Change the script, re-run it, and the entire video re-times itself. Nobody touches a timeline.
Photographed, not recorded	It isn't screen-recording. It calculates what the screen looks like at each moment and photographs it — so the picture can't drift off the audio no matter how slow the computer is.

On cost: both services bill by usage and ElevenLabs is the bigger of the two. Don't quote a dollar figure on camera unless you've checked current pricing first — you don't have a verified number for this build.

The close — the last thing in the video

Nothing before this point mentions the download. This is the only ask, and it comes after the payoff.

Beat 10 · the only CTA

Go take it

The free build pack is in the description — prompts, folder structure, tools, the order to do it in
No email, nothing to sign up for
Then cut. No recap, no goodbye, no subscribe speech.