WorkBuddy Turns Three Prompts into a Finished MP4 in 14 Minutes
An AI agent that can install missing dependencies and fix its own subtitle glyph bugs is ready for small-budget video production, but that 13.7 MB MP4 still failed to play smoothly locally. The automation is real; the last step of human acceptance remains.
WorkBuddy, Tencent's desktop agent that can read and write local files and install its own dependencies, ran a complete short-video production from three natural-language prompts. The agent generated four 3D storyboard images, cloned a voice via edge-tts, burned subtitles, and composited a vertical MP4 – then caught and fixed two subtitle rendering bugs on its own. The whole job took about 14 minutes.
The notable piece is the self-healing: WorkBuddy discovered the system lacked edge-tts and installed it, and later diagnosed that Windows \r\n line endings were causing drawtext to render '口' glyphs in subtitles, then rewrote the script and re-rendered. It also ran per-frame QC on frames 2 and 4 to verify Chinese text like '智能空气炸锅 / 180℃ / 25分钟' was correct.
The final output, a 13.7 MB MP4, wouldn't play smoothly in the local player – stuck at decoding. WorkBuddy's own positioning is as a digital employee for a one-person company, and its integration with the MCP ecosystem means small merchants can hand it the whole video job instead of juggling separate AI chat, image, and editing tools.
The agent's self-diagnosis of Windows \r\n line endings in subtitle text shows automated debugging can now operate at the system level, not just on code syntax.
Running per-frame QC before delivery is effectively unit testing for visual output; it trades a few seconds of rendering for preventing visible defects.
The final playback failure suggests the pipeline optimizes for generation and file delivery, not for playback compatibility; the built-in player already warned about the codec, so the agent could have flagged it earlier.