ContentHQ

ContentHQ is the system behind my videos. It finds what is worth talking about, helps me write the script, and renders the finished reel, all without a video editor open. I built it because most of the time that goes into a short video is mechanical, and mechanical work is what code is for.

Three parts

Pulse: what is trending right now

A personal dashboard that pulls about 27 news feeds and 6 “buzz” sources, groups the items into topics, and scores them. I read it as swipeable cards and go from a topic to a script in one tap.

• Items are clustered by headline keyword overlap inside a 48-hour window.
• The score mixes recency (a 12-hour half-life), how many independent outlets carry the story, and how much buzz it has.
• Each topic stores titles, links and a 60-word summary written by a small Claude model. Article text is never stored or shown.
• Built with Next.js, Tailwind and Postgres, with scheduled jobs for ingesting and pruning. It sits behind a password with a signed session cookie and throttled logins.

The reel engine: one JSON file in, one finished reel out

I write a single reel.json with the scenes. The engine does the rest on my Mac, with no API calls and no AI tokens per video:

1. Voice. Each scene is spoken by Kokoro, a local text-to-speech model.
2. Word timing. Whisper listens to the voice track, so every caption word gets its exact start time.
3. Timeline. Markup in the text becomes caption blocks, and cues like "@fourteen" fire a widget on that spoken word.
4. Frames. A web page draws any moment of the video from seek(t), and headless Chrome captures every frame.
5. Audio and encode. Voice, a quiet music bed and sparse sound effects are mixed and normalised, then encoded to a 1080×1920 H.264 file.

A 55-second reel renders in about two to three minutes. Before rendering, a 20-second “stills” pass gives me contact sheets to check.

Stream to shorts

A five-step pipeline that takes one livestream link and produces one to three reviewed shorts. Every step writes a file the next step reads, and every step ends with a review in ClipsHQ.

Why it’s built this way

• Files between steps. A fix later on means editing one file, never redoing the steps before it.
• Local where it’s cheap. Voice, timing and rendering cost nothing per video, so I can iterate freely.
• A human gate at every step. AI proposes the clips, the plan and the edit. I approve, reject or nudge each one.

What I learned

The video style lives in three files: a stylesheet, a caption engine and a set of widgets. Treating video as a web page that can draw any timestamp made it testable. I can open a preview in a browser, change a colour, and see the reel update before spending minutes on a render.