Two to three hours for thirty seconds
In July I made every clip with me in the loop. The typical interview ran over an hour. The part worth posting lasted about thirty seconds.
Finding that moment was only the start. Then came round after round of render, watch, spot the problem, fix it, render again. Each clip took 2 to 3 hours of my hands-on time, spread over a day or more. I planned three posts a day and averaged less than one. Every step waited on me.
An engine that works the night shift
I built the system with Claude. At 2 a.m. every night, it starts on its own:
- 01
Find
Checks a watch list of 200+ podcasts, channels and conference feeds, picks up to 8 promising interviews and scores every moment for hook, relatability and novelty.
- 02
Create
Cuts the best moments into clips and writes the hooks and captions.
- 03
Check
A frame check and four AI judges, each playing a different kind of viewer, screen every clip, so weak ones never reach me.
- 04
I approve
By morning the best are waiting on a review page. I pick the ones worth posting, and the engine loads them into TikTok as drafts for me to publish. My edits to hooks and captions feed back into how the engine writes the next ones.
More than twice the posts, a fraction of the time
The engine didn't just speed up my old process. It replaced the parts that never needed me.
| Measure | Before · July | Now · September |
|---|---|---|
| Finding what to clip | 1 at a timeI found each interview myself, typically longer than an hour, for about 30 seconds of clip | 200+sources on the watch list, checked overnight. The AI mines up to 8 talks while I sleep |
| My time per post | 2–3 hrsof hands-on work per clip, spread over a day or more of render, watch, fix | ~11 minto watch the clip, pick a caption and post it |
| Posts per week | 6in my best weeks of July | 142.3× more |
| Ready each morning | 0clips. Nothing existed until I made it | ~6finished clips ready to post (3 to 8 a night) |
Data as of Sept 21. Hours back are estimated from earlier git commit logs.
Two lessons I bring to your build
Good help checks its own work.
A good assistant proofreads a letter before it lands on your desk. My engine does the same: four AI judges, each thinking like a different kind of viewer, screen every clip, and the weak ones never reach me. When I build for you, every draft is checked against your rules first, so your time goes to deciding, not fixing.
It should learn from you, not run without you.
Every batch I review sharpens how the engine writes hooks and captions, but nothing posts without my OK. In a business where one wrong message can cost a client, that is the design, not a gap.
The media workflow isn’t the point. The pattern is.
AI finds the work → drafts it → checks it → you approve → your edits teach the next run.
The same shape, in my engine and three local businesses:
- My engine. 200+ shows and feeds → up to 8 promising interviews → clips, hooks and captions → a frame check and four AI judges → I post the best ones each morning.
- Financial planner. The day’s meeting notes → every promise and to-do → a follow-up email and CRM notes for each meeting → checked against the client file and compliance rules → you send the follow-ups before your first meeting.
- Real-estate agent. The whole contact list → past clients due for a check-in → a personal note to each → checked against your do-not-contact list and CRM → you send the ones you like over coffee.
- Insurance agency. The next 60 days of renewals → the big premium jumps → a heads-up note and options for each client → checked against the policy file → you reach every client before the renewal bill does.
In every one, the AI continuously improves based on your feedback, and your results will consequently improve.