Twelve hours to a channel that runs itself

The bet was simple to state. In one evening, stand up a media brand from nothing (accounts, logo, voice, animation, editorial line) and leave behind a machine that publishes two videos a day on three platforms without me. Not a prototype that needs someone to nurse it through each run, but something that retries, refreshes its own tokens, and holds an editorial line when nobody is watching.

Hours 0 to 2: the spine. While I created the accounts in a browser, a research agent compared six programmatic-video stacks and came back with Remotion, mostly because videos are React components there and headless renders come out the same every time. By midnight the pipeline existed end to end: an episode JSON in, a finished 1080x1920 short out, with a Hume Octave voiceover, whisper.cpp word timestamps, and TikTok-style burned captions.

Hours 2 to 4: the brand. The logo went through roughly ten iterations of a design loop where the model renders a frame, actually looks at it, critiques what it sees, and fixes it. I rejected every blob-shaped brain until an anatomical side profile read correctly: solid left hemisphere, dotted outline completing the right, a neural network living inside. The rejects still sit in the brand kit.

Hours 4 to 6: the operator. The daily engine is not a script. It is a 170-line runbook that a headless Claude Opus session re-reads every morning at 06:57: pick the next curriculum topic, verify the day's biggest AI story across two sources, write both scripts and strip them of AI cadence, choose two different natural voices with acting instructions for tone, render, extract ten frames and inspect them with vision, fix what looks wrong, then publish. The first fully unsupervised run took 8 minutes and both videos were publishable.

Hours 6 to 8: taste, enforced by experiment. I ran the same morning three ways on identical instructions: Sonnet from scratch, Opus editing Sonnet, Opus from scratch. Prose quality tied. Judgment did not: clean Opus caught that Sonnet had attributed a story to an outlet that never ran it, and it declined to burn upcoming curriculum topics early. The model tier was decided by that test, not by vibes. Every taste note from those reviews (captions in the lower third, no announcer voices, no em dashes, ever) became a permanent runbook rule.

Hours 8 to 12: the rails. YouTube's documented forced-private policy for unverified API projects turned out not to trigger; I tested it with a throwaway unlisted upload before trusting it. Instagram's modern API refuses direct byte uploads and demands a public URL, so each video now spends about three minutes as a GitHub release asset before the container is published and the asset is deleted. Facebook granted every permission and then returned an empty page list anyway; the fix is fetching the page token by direct ID. Four landmines in one evening, and each one is now a line in the runbook so the next run does not rediscover it.

The part I would defend in a design review is the runbook-as-product idea. The pipeline code is ordinary. The asset is the accumulated judgment in one prompt file: voice rules, curriculum logic, QC checklists, failure policies (retry once, skip, report, never double-post), and every scar from launch week. Corrections compound. The machine gets more like its editor every day it runs.

The build itself ran on a cost gradient, and so does the machine it left behind. A frontier model sat in the architect seat, but it delegated relentlessly: a Haiku scout swept codebases before decisions, Sonnet subagents did the bulk work (a six-stack video-framework comparison, wiring the QC step, integrating this very page), and Opus took anything that needed judgment. The daily automation inherits the same economics. The expensive model runs for about eight minutes a morning, at the points where a wrong call would end up in a published video, and the cheaper models do the sweeping.

What I would claim is narrow. An agent can hold real production responsibility (research, editorial judgment, rendering, publishing, token refresh) when the system around it supplies bounded permissions, a way to check its own output, and somewhere for human corrections to accumulate.

What I would not claim yet is durability. The whole thing hangs off one launchd timer on one laptop, so a machine that stays asleep is a missed morning. The QC pass looks at frames, which means it catches a broken render and not a boring script. The Instagram path leans on GitHub release assets as a staging CDN, which I took because the API left no other route, not because I wanted a second service in the publish path. And the track record is still only as long as launch week.