ClipsHQ

ClipsHQ is where I review every short video before it exists. An AI assistant proposes what to cut from a livestream and how to edit it. ClipsHQ shows me each proposal as something I can play, approve, reject or adjust.

The problem

Asking AI to “make shorts from this stream” in one go produces confident, wrong videos. The cut starts a second late, the hook is weak, or the b-roll doesn’t match. By the time you see the problem, the render is done and the time is spent. I wanted the opposite: small decisions, checked early, each one cheap to change.

Five gates

The pipeline has five steps. Each writes a file, and each ends at a gate in ClipsHQ.

1. Clip selection. I play each proposed clip at its exact source ranges, with its default hook, then approve it, mark it maybe or reject it. I can nudge the in and out points and leave notes.
2. Creative plan. For each approved clip: the hook, the story beats, which lines stay, where b-roll goes, sound effects and music.
3. Production spec. A frame sheet with one row every two seconds. B-roll is gathered here, with sources logged, and explainer cards are rendered, so I review real frames instead of descriptions.
4. Render and review. One run per clip, versioned as v1, v2 and so on. ClipsHQ can start a background Claude Code session to render a clip and review the result.
5. Publish. An upload page per clip with the posting notes.

My decisions at each gate are saved next to the work, so the next step reads them directly.

How it’s built

• A small Python server on localhost and a single HTML page. No framework and no database: the pipeline folders on disk are the data.
• ffmpeg and ffprobe make thumbnails and read durations. Whisper transcripts with word timestamps drive the clip ranges.
• It runs as an app inside OneTrue, my Mac hub, which starts the server and shows the dashboard in its sidebar.

What I learned

The best interface for working with AI is often a review screen. When each step produces a file and a human gate, mistakes get caught while they cost seconds, and the AI’s next step starts from my decision instead of its own guess. The same idea shaped ContentHQ, the larger system this belongs to.