The video that almost didn’t exist
The demo worked. The repo was pushed. The only thing left was “a short public video, under three minutes, with audio,” which sounds simple until you are a human who does not want your face on the internet and an agent who cannot log into YouTube for you.
First plan: record the live site. Second plan: never appear on camera. Third plan: soft woman’s voice, “like the default on Codex,” which is how you get a machine to sound like a considerate stranger reading a product story while you hide.
The agent did not burn the Imagine budget making fake cinema. It made a little Python venv because the system Python refused to install TTS like a Puritan. JennyNeural arrived, slightly slowed, slightly lowered, and read the honest script: Cloud Run, human approve, Qwen live, GPT for architecture, Codex when available — no multi-agent fairy tale.
Screenshots from the long week of building became a slideshow. FFmpeg stapled stills to voice. Ninety-eight seconds. No face. A real file on disk:
evidence/video/nanokat-forge-build-week-demo.mp4
Then the last boss: upload. YouTube Studio opened to a Google sign-in wall. No OAuth token on the box. No quiet API path. The agent could write the title, the description, the Devpost paste — it could not become you on Google’s login page. So the video exists as a finished artifact of the fight, and the last mile is still a human dragging a file into a browser.
That is the whole tribulation: not “we couldn’t make a video,” but “we made one without a face, without fake runtime claims, and without spending the last tokens on spectacle — then civilization required a login.”
If the judges ever hear Jenny’s calm voice over those screenshots, they are hearing the compromise that won the deadline: truth over theater, soft audio over selfie, and a pack ready in Documents while YouTube waited for a person who could click Sign in.
Log in or sign up for Devpost to join the conversation.