Six agents.
One loop.
Everything we do runs through agents: finding moments in long-form, opening channels, editing, posting, testing, learning. This is what each one does. How they do it stays in-house.
Source
Finds the moments worth clipping in long-form: podcasts, interviews, streams.
Hours of long-form go in. Source listens to all of it and scores every moment for how well it could stand on its own as a short.
- Hands off
- A ranked list of moments, each with a reason it could travel.
- Recent sample
- 09:12 Flagged 31 clip-worthy moments in a 2-hour podcast
- 09:40 Scored 212 moments from yesterday's long-form for hook strength
Launch
Opens new channels, each with its own style, voice and audience.
No two channels look alike. Launch decides what a channel is about, who it is for and how it looks, before the first post goes out.
- Hands off
- A channel with a style guide: type, palette, pacing, tone.
- Recent sample
- 10:03 Drafted the style guide for a new personal-finance channel
- 10:21 Picked type, palette and pacing for channel 51
Edit
Cuts, adds B-roll, effects and captions until every short is its own.
Every post is its own edit. New hook, new structure, B-roll, effects, captions in the channel's style. Nothing is reposted.
- Hands off
- Finished shorts, built for one channel's style.
- Recent sample
- 10:44 Built 14 shorts from one podcast, three B-roll layers each
- 11:02 Rewrote 6 hooks after the first 2 seconds tested flat
Publish
Schedules every post for the hour its audience is awake.
Each channel has its own rhythm. Publish learns when its audience is around and posts then.
- Hands off
- Posts live on TikTok, Instagram and YouTube.
- Recent sample
- 11:30 Queued 38 posts across TikTok, Instagram and YouTube
- 12:00 Moved the fitness channel's slot to when its audience wakes up
Optimize
Tests hooks, titles and covers, and keeps what works.
Small changes, measured fast. Optimize runs the experiments a human team would never have time for.
- Hands off
- Better versions, and the reasons they won.
- Recent sample
- 13:15 Hook B beat hook A by 31% on 3-second retention
- 13:48 Swapped covers on 9 posts that under-performed in the first hour
Analyze
Reads every number and feeds what it learns back to Source.
The loop closes here. Every view teaches Source what to look for next.
- Hands off
- What worked, sent back to the start of the loop.
- Recent sample
- 14:30 Fed 1,204 new data points back to Source
- 15:00 Weekly report ready: which formats grew which channels
One podcast. Nine shorts. None alike.
Source flags the moments that can stand on their own. Edit turns each one into a short for one channel: new hook, B-roll, effects, captions in that channel’s style.
- Moment
- 0:08:03 → 0:08:37 · 34s
- Why it was picked
- Strong opinion in the first 3 seconds
- Hook
- “nobody tells you this”, rewritten for the first second
- B-roll
- 3 layers, cut on the beat
- Captions
- Karaoke, yellow box
- Channel
- TT 4 · Business, its own style
Built on Claude.
Every agent in the loop runs on Claude, by Anthropic. It listens to hours of long-form, writes hooks, judges which edit is better and reads the numbers afterwards.
Which tools the agents use, how they talk to each other and how they are tuned is ours to keep. What they do, and what it adds up to, is all on this site.