ALAN NICHOL

Alan makes marketing videos for Rasa, the AI-agent platform he co-founded, where the on-screen host is a deepfake of himself. He records a couple of minutes of audio between meetings, and an agent generates the rest: Claude writes the Remotion code (a JavaScript library for video as code) that animates the text and graphics, and HeyGen turns his recorded voice into a talking-head avatar of him, only the audio is real. The skill carries the one thing Claude has none of: his visual taste. Work that used to take a creative agency weeks, done solo from the terminal.

▶ WATCH SEGMENT $ REMOTION-VIDEO SKILL

100Xcheaper than a creative-agency cut, so these videos get made at all NO GUIthe whole video built from the command line, never touched one AI MEonly the voice is real, the talking head is a generated avatar WHISPERword-level timestamps, so on-screen text lands on the spoken word ONEone focal point on screen at a time, never busy animation MSa jump cut and dead energy are milliseconds apart

this episode is also a blog post: The Agentic Software Factory and a podcast: Agentic Engineering and the Lost Art of Verification

EP 03 · ALAN NICHOL · making product videos with a coding agent, live on stream

"Everything that can be code will be code."

His operating thesis. The productivity gain from making something code, then editing it with a coding agent, is so large that "nothing survives that." Video became interesting to him the moment Remotion made it code-shaped. 02:46:28

THE SKILL IS ALL TASTE

claude writes the remotion fine. the skill carries what it can't write

The skill holds none of Remotion's syntax. Claude already writes it fluently. What it holds is everything Claude lacks: layout modes, canvas safe areas, minimum font sizes, easing functions, animated text that never just repeats the narration. It is a living document, rewritten every time Alan learns something new about what he wants out of a video.

He built it in a domain where, by his own account, he is a novice without the vocabulary. He made vague subjective requests like "make this a bit more high energy," watched Claude turn them into easing and transitions and timing, and wrote the working vocabulary back into the file. The gap that remains is his own: "if I knew more about videography, I would be able to give much more precise instructions."

Alan Nichol's remotion-video skill open in a browser tab, showing the conceptual design rules as text — Alan opens the skill in a browser and walks its conceptual design rules. "Claude needs no help whatsoever writing Remotion code," so every line here is taste, not syntax. [02:51:08]

"Prompt and pray."

Alan's name for the trial-and-error loop with no systematic path from a bad output to the one you want. He puts it on stickers and leaves them on conference tables. It is the same throwing-stuff-at-the-wall he remembers from deep learning in 2014. The skill, and the render-frames check, are his way out of it. 02:42:06

AUDIO IN, VIDEO OUT

a few minutes of recorded voice, and the agent generates the rest

Record a few minutes of audio"I just record some audio into my computer, quickly in a phone booth or something when I have a few minutes." No lighting, no re-records, no calendar time. 02:49:57

Claude writes the RemotionAnimated text, graphics, and the animation of the graphics, all defined as code. "Claude needs no help whatsoever writing Remotion code. So it will just do it." 02:49:05

Generate the talking head with HeyGenThe on-screen Alan is a deepfake avatar of him, built from his real recorded voice. HeyGen's V4 model snapped back to an unsettling base face on pauses, so he trims around dead air, sometimes mid-word. 03:09:24

Align text to the audio with WhisperWord-level timestamps let Claude pop text or an animation as Alan says it, or just ahead of it: "you want to do it just a hair before you say it." That makes the video feel like it is leading. 03:11:23

Render frames and inspect"You really need that verification loop. It's 100 times faster and more effective if you give it a way to inspect the output." Claude takes no video input, so the skill makes it render individual frames and look. It catches some problems; Alan still spells out the obvious ones. 03:00:16

Alan Nichol playing a generated Rasa product video on screen share, with speaker tiles beside it — The video Alan plays on stream is about emergent personalization; the orange-backed talking head is his generated avatar, the voice his real audio. "No one's going to watch those and think they were done by Ridley Scott, but the fact that it's me solo and I'm able to produce them." [02:50:05]

TWO VIDEO SKILLS, ONE SHIPPED

the polished one, and the vertical-video experiment he's openly paying for help on

skill

remotion-video

Encodes his visual judgment for programmatic video, so Claude turns a few minutes of recorded audio into a finished explainer. The write-up captures the design rules he walks on stream. He keeps rewriting it as he learns, so read it and build your own.

skill · live demo

short-form vertical video

A second skill that re-cuts the raw horizontal video into vertical short-form: full-screen talking head, animated subtitles, illustrations over the face, aggressive jump cuts. Not good enough to run yet. "I will happily pay you to help me fix this."

"I thought I was a power user of this stuff, and then I realized I am a little baby. I don't know anything."

Alan, after watching the other Episode 3 guests. He treats the not-knowing as the point: the video work puts him back in a beginner's mindset in a domain he understands nothing about, and he can still ship. 02:40:30

THE STACK

what turns a few minutes of audio into a finished video

Claudewrites the code, has none of the taste Remotionvideo as code, in JavaScript HeyGenthe talking-head avatar from his voice Whisperword-level timestamps to align the text Rasahis platform's bet: "vibe code is in" 2026.rasa.comthe videos, made exactly this way