How an episode is made
An episode is a folder: a manifest, the scene's code, the narration, the rendered video, its 90-second edit, a thumbnail and a status record. Everything that happens to it is written there, so any episode can be audited, re-rendered or re-posted.
1. Choosing what to make#
The topic planner reads the channel's niche, its topic bank, what it has already posted, what performed (from the R&D wing) and a variety rule, and proposes a subject with a hook and an angle: a surprise, a consequence the viewer can see, a number that can be checked. Every topic in the bank carries its angle, the surprising facts, the consequences to show on screen and a 'wow' shot; premises everybody already knows are retired, and so is any topic that would repeat a video already made.
2. Writing the episode#
The episode agent — Claude Fable 5.2 since 10 November, working in a sandbox on the GPU box — receives a brief: the topic, the channel's kit documentation, its rulebook (a dozen rules distilled from that channel's past failures), the pacing rules and the quality bar. It writes the scene, the narration script and its on-screen labels, renders stills to look at its own work, fixes what it sees, and iterates until it is satisfied. If a model or a cloud account refuses a call, the call moves down a chain — Fable 5.2, Fable 5.1, Opus 5.5, Opus 5, Opus 4.8 — and across the cloud accounts that may serve it, so a model outage never stops a batch. The scene is either code for the channel's WebGL2 renderer or, for a Blender episode, a Python build of a Blender scene (next chapter).
3. Rendering on a GPU#
A code scene is rendered by headless Chrome with WebGL2 on an NVIDIA L4, frame by frame at 1080×1920, and encoded with ffmpeg. Render parity across machines is proven pixel for pixel: the same reference stills come out byte-identical on every box of the fleet. A Blender scene is rendered by Blender 4.5 LTS on the same L4s, headless, in one and a half to five GPU-hours.
4. Narration, captions and the mix#
The narration is written for the ear — numbers spoken as words, units as a narrator would say them — and read by a Google Chirp 3 HD voice chosen per channel. A speech recogniser on the GPU then hears every line and times each word, and the captions are drawn from those times: the spoken words themselves, each one lit as it is said (the chapter 'Captions that are the voice'). Music and effects are mixed under the voice, and the result is levelled to streaming loudness (−14 LUFS, true peak below −1 dBTP). A second, shorter cut of 55 to 89 seconds is made for Facebook and Instagram, whose Reels limit is 90 seconds — its joins snapped into the pauses between words, and the edit must tell the story on its own.
5. The quality gates#
| Gate | What it checks | What happens on a fail |
|---|---|---|
| Critic | an independent review of a contact sheet, the captions and the measured motion: visual quality, readability, factual accuracy and policy; 8/10 or more and no blocking issue | revised with the notes, else held |
| Entertainment | part of the critic's review: would a viewer keep watching — consequences shown, motion, surprise (7/10 or more) | revised, else held |
| Video and voice | 1080×1920 at 30 fps, 120–175 s, H.264 and AAC; the channel's narrator really in the soundtrack | re-rendered, else held |
| Captions | every spoken word on screen, in order, at the moment it is said; what was drawn matches the word log, frame by frame | drawn again, else held |
| Caption sync | the finished soundtrack heard again: how far each caption is from its word (median, 95th percentile, share of words matched) | measured on every episode; enforced once its thresholds are final |
| Text and units | no text outside the one text band; SI symbols on screen, units spoken in words | fixed, else sent back |
| Picture and page | no black or frozen segment; no JavaScript or shader error in re-rendered stills | re-rendered, else held |
| Flash | photosensitive safety (WCAG 2.3.1): no more than three flashes in any second | held |
| Loudness | integrated −15.5 to −12.5 LUFS, true peak at most −0.5 dBFS | re-mixed |
| The 90-second edit | a viewer model reads the edit's narration alone and must follow it (8/10); else a story editor picks better lines and the edit is cut again | Facebook and Instagram wait for a short version |
6. Library, archive, publishing#
A passing episode enters its channel's library as ready. Within minutes the portal's content sync copies every file into the media archive on Google Cloud Storage, where it is versioned and kept, and every hour the whole archive is copied into a second bucket; the control host removes its own copy only once the archive holds the video at its exact size. From the library, slots take the oldest episode each platform has not had, so every platform of a channel posts the same episodes in the same order — a Blender slot takes the oldest Blender episode, a code slot the oldest code one.
Lessons that make it better every day#
Every episode that needed a fix, failed a gate or was rejected by a reviewer leaves a lesson. When a channel has gathered a few new ones, a 'doctor' model distils them into that channel's rulebook — at most a dozen sharp rules — which is put at the top of every later brief. Mistakes are therefore made once. The same mechanism carries across channels when a lesson is general.