Scene 1 · The Hook
“How does an AI actually understand words? It doesn't read letters—it turns features into direction arrows on a giant map!”
Lower third: AI maps concepts as direction vectors.
▶️ Watch the Short on YouTube · youtube.com/shorts/8ud3AaGKtP0
📺 YouTube upload page (title, description, thumbnails) →
A 48-second YouTube Short that explains how AI understands words as vectors — with a horse. Left: the output. Below: the input from Google Flow and how it was assembled.
short_1080x1920.mp4 — 1080×1920, 24 fps, H.264/AAC, 49.5 s, 18 MB.
Six clips stitched, ElevenLabs voice-over (via fal.ai), animated word-by-word captions, a 2-second Skool end screen, and a crossfade from the end screen back to the first frame so it loops.
🚀 Join the community: Delivery Pilot on Skool
Each clip is 8 s, vertical 9:16, with its lower-third card already burned in. The voice line below each is the ElevenLabs take used in the export.
“How does an AI actually understand words? It doesn't read letters—it turns features into direction arrows on a giant map!”
Lower third: AI maps concepts as direction vectors.
“Think of this horse as a base vector. Every single concept sits at its own specific spot on the coordinate grid.”
Lower third: Base Vector Node: Horse
“Add the “Wings” direction vector to the Horse... and the AI moves directly to the coordinate for Pegasus!”
Lower third: Horse + Wings = Pegasus
“Add the “Horn” direction vector instead... and the coordinate shifts straight to Unicorn!”
Lower third: Horse + Horn = Unicorn
“Which means... Pegasus, minus Wings, plus Horn, equals Unicorn! To an AI, concept logic is just basic geometry.”
Lower third: Pegasus - Wings + Horn = Unicorn
“Token 9246 is hiding inside! But... how do these vectors talk to each other? How DOES an AI understand concepts?”
Lower third: Token 9246 | Seamless Loop
Full written record: docs/input/google-flow-brief.md.
Scene 6 zooms into the unicorn horn where Token 9246 ignites, then returns to the dark star-field of scene 1. The last question flows straight into the first line, so the Short loops.
You started in Google Flow (idea, storyboard, six clips, voice-over script) and moved to Claude Code to build, verify and publish. Source: docs/workflow.md.
flowchart TD
subgraph FLOW["Phase 1 · Google Flow"]
A["Idea: how does AI understand words?"] --> B["King - Man + Woman = Queen"]
B --> C["Simplify for a 4-year-old: Kitty + Wings"]
C --> D["Age up: Horse + Wings = Pegasus, Horse + Horn = Unicorn"]
D --> E["Storyboard images: 6 panels, 9:16, lower thirds"]
E --> F["Script and ElevenLabs voice-over text"]
F --> G["6 video clips, 8 s each"]
end
subgraph CLAUDE["Phase 2 · Claude Code with superpowers"]
H["Hand-off: 6 clips, scene list, VO script"] --> I["Brainstorm: pick Remotion"]
I --> J["Design doc"]
J --> K["Implementation plan"]
K --> L["Execute test-first, commit to main"]
end
subgraph SHIP["Phase 3 · Ship"]
M["Export MP4"] --> N["GitHub Pages site"]
N --> O["YouTube page: title, description, thumbnails, pinned comment"]
end
G -->|"download to Downloads"| H
L --> M
flowchart TD CL["assets/clips: 6 Flow clips"] --> ING["npm run ingest"] SC["src/script.ts: scene lines"] --> TTS["npm run tts: fal.ai ElevenLabs"] TTS --> VO["assets/vo: scene1-6.mp3"] VO --> AL["npm run align: fal.ai Whisper"] SC --> AL AL --> CAP["assets/captions.json: word timings on script words"] ING --> RM["Remotion composition"] VO --> RM CAP --> RM LOGO["skool-logo.png"] --> RM RM --> MP4["out/short_1080x1920.mp4"] MP4 --> VER["verify-export: format, 49.5 s, under 25 MB, audible"] VER --> GIT["git commit and push to main"] GIT --> PG["GitHub Pages workflow: static.yml"] TH["scripts/thumbnails.mjs: fal.ai nano-banana-2"] --> YT["youtube.html"] PG --> YT
sequenceDiagram autonumber participant You participant Flow as Google Flow participant CC as Claude Code participant Fal as fal.ai participant GH as GitHub You->>Flow: Idea, then simplified, then Horse/Pegasus/Unicorn Flow-->>You: Storyboard, script, 6 clips You->>CC: 6 newest videos, scene list, VO script, fal key CC->>You: Approaches (Remotion chosen), design, plan CC->>Fal: ElevenLabs TTS per scene Fal-->>CC: Voice-over mp3 files CC->>Fal: Whisper for word timings Fal-->>CC: Timestamps mapped onto script words CC->>CC: Remotion render and verify-export You->>CC: Add CTA, move captions, thumbnails, pinned comment CC->>Fal: nano-banana-2 vertical thumbnails CC->>GH: Commit and push to main GH-->>You: Pages site and YouTube page
flowchart TD B["brainstorming: approaches A / B / C"] --> D["design doc"] D --> W["writing-plans"] W --> E["executing-plans"] E --> T["test-first per task"] T --> V["verification before completion"] V --> M["trunk-based commit to main"]
The whole video is code in this repo, so anyone can re-run it. Here is the path from idea to MP4:
npm run ingest copies the six newest clips from Downloads into assets/clips/scene1…6.mp4 in story order.npm run tts sends each scene’s line to ElevenLabs through fal.ai (voice Adam, stability 0.4, similarity 0.8, style 0.15). Any line over 7.5 s is regenerated slightly faster so it never outruns its clip.npm run align transcribes each voice file for word timestamps, then maps them onto your script’s own words, so “Pegasus” and “9246” are always spelled right.npm run build:video renders out/short_1080x1920.mp4; node scripts/verify-export.mjs confirms format, 49.5 s length, under 25 MB, and audible voice.main; GitHub Pages publishes this page, the clips and the final video.npm install
cp .env.example .env # add FAL_KEY
npm run build:video # ingest → tts → align → render
npm run studio # preview in Remotion Studio
Built test-first with the superpowers skills: design → plan → execution.