The system behind my videos
Every part of how my videos get written, edited and measured. Each part links to the episode that built it, and the editing parts are the series I'm filming now.
Content creation
How an idea becomes a script, two LinkedIn posts and a newsletter issue.
- built
Ideas from my voice chats land in a Drive folder as files. A pickup turns each file into a Fizzy card and marks the file done.
- built
Two Fizzy boards: TB Content for ideas and episodes, Pipeline build for the system itself. Every next step is a card here, never a to-do in a file.
- built
My voice file, my bio and the channel strategy. Every skill reads them first, so a draft sounds like me before I edit it.
- built
One argument sentence, the brief, title and thumbnail, five hooks, then the draft with four markup tags. Writes the brief, the script and the teleprompter copy.
- built
A separate agent reviews the draft against the script skill's rules and checks every named fact against a source.
- built
Reads the script's visual tags and proposes a visual for each gap, marked as a file from the folder, a screen recording, a motion unit or camera only.
- built
I strike lines in the shotlist and rule on the evaluator's findings. The last thing a human touches before the overnight run.
- built
Two posts per episode, lifted from the script's sections, each with its own five hooks.
- built
The System Layer. Signup on the site, double opt-in and the drip in the hub engine. Each issue carries the episode's full count and its first rule.
- built
tadejbokan.com: the signup page, the short links that tag every published link by episode and placement, and this map.
Video editing
The series: the machine edits each episode overnight, I count what it got wrong, and the rules grow from the count.
- coming
One teleprompter take from the stripped script; when I stumble, I re-read the sentence. OBS records face and screen to separate files from a locked profile.
- coming
A word-level transcript with a timestamp on every spoken word, run on my PC.
- planned
Reads the raw transcript against the script: restarts, English slips, filler count, and the sentence shapes I stumble on, which go back to the evaluator so the next script avoids them.
- coming
Aligns the transcript to the script, keeps the last take of every sentence and drops the rest. A sentence it cannot match stays in full and is flagged; nothing is dropped silently.
- coming
The first thing I built in: level to -14 LUFS, no clipping, in sync. An episode below this floor is re-run, never published.
- coming
For each shotlist line: a file from the episode folder, a screen recording at the matching moment, or a motion unit. Nothing from the internet, nothing shot specially.
- built
Dark ground, one signal colour, two typefaces, type sizes as a share of frame height, and the defaults that are banned. Baseline, not a rule.
- coming
A library of unit types rendered from the script's data files: text pop, stat card, diagram, list. Each type is built once at high effort through a three-round critique at phone size, then reused.
- coming
Applies the cuts, places the units at their anchors, lays the music bed, writes the chapters. The output is the draft I publish untouched.
- coming
The same edit as an FCP7 XML timeline for Premiere or Resolve. I open it only to count what I would change, never to publish from it.
- coming
Checks for black frames, drift, loudness, overlays outside the frame or on my face, and lists what the run skipped. Morning delivery is the draft plus this list.
- coming
I watch the draft with the QC list open and write timestamped notes. Every note that is a rule, not a one-off, goes into the rules file.
- coming
The edit skill's rules, grown only from my notes on drafts. The count on screen in every episode starts at zero.
- planned
Rewrites the series overview, the per-episode log and the rules ledger from the episode folders, so every number in the series files comes from an episode file.
Analytics and learning
Which content brings newsletter signups, which brings buyers, and what I make next because of it.
- planned
One log per episode: tokens and cost, what the machine got wrong, my count of changes, how I delivered. The next script is written from these numbers.
- planned
Per episode: how many changes I would make, of what kind, and how many minutes marking them took. Whether it falls toward zero by episode ten decides between the three endings.
- built
Every published link carries the episode and the placement it was published in, so each newsletter signup is attributed to the video, post or issue that brought the visitor.
- planned
Weekly pulls of the raw numbers from YouTube, LinkedIn, the newsletter and the site into one table, so the history exists before the analyst does.
- planned
Two modes. On demand: I give Hermes a YouTube link; it reads the transcript and the comments weighted by likes and replies, compares what the video covered with what viewers asked for and did not get, and returns the five gaps with the comments that prove them and the angle to take. On a schedule: it runs the same analysis on my own videos, from my scripts in the repo, my channel's comments and the analytics, and writes each gap worth a video to the TB Content board as an idea card.
- planned
The content analyst. From the logs, the collectors and the attribution it answers: which videos, posts and issues brought the most newsletter signups, which brought buyers, which parts of the system cost the most, and what I should make next. It runs the gap analysis on my own videos on a schedule and puts the video ideas it finds on the TB Content board.
- planned
One button on every node of this map: I want this video. Demand per unbuilt part, counted and shown above a threshold.
Newsletter
The System Layer
Each issue carries the episode's full count and the first rule it produced.