Skill Autoshorts

Skill Autoshorts is an agent skill that turns long vertical videos into short clips for TikTok, Instagram Reels, and YouTube Shorts, with a human approval step in the middle. You forward a recording to your agent in chat; the agent saves it to an input folder. Once a day the skill picks the newest unprocessed video, chooses moments worth clipping, cuts them, overlays a hook text, and posts a candidates table back to the same chat. You reply with the clip IDs to publish, and the skill schedules them through the Upload-Post API.

Clip selection combines two models. Whisper (medium, run locally) produces a word-level transcript that acts as the clock, so every cut starts and ends on a word boundary. Gemini Flash receives the full video file plus the transcript through the Files API, so it can react to visual cues such as laughter, gestures, and scene changes, and the prompt requires it to snap clip boundaries to Whisper timestamps. ffmpeg performs the cut and Pillow renders the hook overlay. Input footage is expected to be 9:16 (ideally 1080x1920) with subtitles already burned in; the skill does not crop, reformat, or add subtitles.

The skill targets agent harnesses such as Hermes and OpenClaw running on a VPS: the harness fires the daily cron, relays the candidates to your messenger, and pipes your reply back. It also runs as a standalone CLI with pick, transcribe, analyze, extract, hook, publish, and mark-processed subcommands. A sha256-keyed state/processed.json is the only state between runs and prevents a video being clipped twice.

A feedback loop feeds engagement data back into selection. Each published clip's context is logged; a weekly learn command pulls platform metrics from Upload-Post, ranks clips by a composite z-score (0.6 x views + 0.4 x engagement rate), and asks Gemini to rewrite a HOT.md pattern file that is prepended to every subsequent analyze prompt. Clips younger than 7 days or older than 90 are excluded, and runs with fewer than about five winners and five losers skip synthesis. Every run writes an audit file and backs up the previous HOT.md.