An end-to-end AI-video pipeline, proven unattended.

12

videos auto-posted over 4 days before stalling

£0

extra generation cost per clip vs ~$2.42 on a paid API

45

credits per clip against a 1,200/month ceiling

4+

video models tested before locking Seedance 2.0

/ the challenge /

The real target was behavioural: ship on a schedule.

The genuine aim was to ship something on a schedule instead of polishing forever. That surfaced a real backend problem — the convenient video model isn’t cleanly callable from an unattended cron, and paying for a separate video API was off the table — plus authenticity and child-safety risk in the baby-reaction niche.

A scheduled local Claude session fires

The theme picker advances the round-robin and returns the next locked prompt

Claude generates a 10s clip via the video connector (or ffmpeg stitches library clips)

Claude writes the title, description and #Shorts tags in-session

The uploader posts to YouTube, the run is logged, and a failure raises a flag

Verifying that a headless Claude session could reach the video connector and pull the live credit balance — which meant the whole pipeline could run unattended on the existing subscription at zero extra cost, killing the need for a paid video API.

/ tech stack /

Built to be owned.

Proof: The run log — 12 successful automated YouTube posts logged over four days, then failure rows once the queue emptied.