What they did
The founder behind Cybiqon runs a Claude Code agent on a 4-core Ubuntu server with no GPU. Three cron jobs start it each day: a blog explainer, a product Short and a teaching video. For each one the agent picks a topic, with a sub-agent that searches Stack Overflow, GitHub issues and Ask HN for problems developers really have and must cite two links as proof. It writes the script, which then passes a linter, three fresh Claude reviewer sub-agents and a second model from another company called through OpenRouter. It voices the script with open-source cloning models (Pocket TTS for English, IndicF5 for Hindi) trained on the owner's own recordings, builds and renders the frames, and schedules uploads to YouTube and X, checking half an hour later that they went public. Teaching videos go to Telegram for the owner to review first. There are 40 pass or fail gates, eight enforced in code.
How it went
Between 8 and 17 October it made 40 videos, 65 minutes in total, including three long-forms of 6 to 7 minutes. Swapping the paid HeyGen voice clone for offline models took voice costs to zero; the HeyGen credit had once run out mid-pipeline. Each agent run costs $11 to $26 by Claude Code's own estimate, plus about $0.03 for the second-opinion review. Early Reels got 120 to 150 views and 67 percent of viewers left in three seconds, which led to a new target viewer and hook rules. Instagram and Facebook are still posted by hand, and no Hindi video has been published yet.
Worth knowing
Most gates were added the day a video failed in that way, and the owner still found a Short that passed every check taught nothing. The second model caught a script error that three Claude reviewers missed. Reddit blocked the server's data-centre IP. The studio is MIT-licensed and installs as a Claude Code plugin.