EXPLORE IT LIVE
My AI System, on one map
Click any branch to open it. Keep clicking to go deeper. Every branch is live in my own business.
My AI Team
Six AI employees. One closed feedback loop. No humans touching the content pipeline day to day, I still approve every ship and do the sales myself. This is exactly how I run my content business. The roles, the schedules, the handoffs, the memory system. Copy it.
Every solo creator hits the same wall.
You make content. It works. You want more. You burn out. You try automation. But automation can't think. You hire people. They don't care as much as you do. Sound familiar?
The Input
- Your expertise
- Your time
- Your energy
- Your creativity
The Bottleneck
- One human brain
- 24 hours / day max
- Burnout cycles
- Context switching
The Output
- Inconsistent content
- Missed publishing days
- No learning loop
- Same mistakes repeat
The fix isn't more hours. It isn't more software. It's more brains. Each one owning one job, handing work to the next.
6 employees. 1 closed feedback circle.
Each agent does exactly one job. They all read and write to a shared memory file. The last agent in the loop improves the first agent. So the whole team gets smarter every week.
Exactly when each employee clocks in.
Everything is on a cron. No babysitting. I don't lift a finger during the week unless I want to.
Weekly output: 7 carousels · 21 reels across 4 platforms · 1 performance report · 1 self-improvement cycle
The Creator
Writes brand-new carousel posts every Monday morning. Hook, copy, slide design, caption. All of it, from scratch, based on what worked last week.
INPUT
- Last week's performance report
- Memory file (top hook patterns)
- Optimizer's topic recommendations
- Current queue depth
- Sam's NotebookLM (for real stories)
PROCESS
- Pick 3 topics from data
- Research in NotebookLM
- Fact-check every claim
- Write slides + caption
- Render HTML carousel
- Score 80+ or reject
OUTPUT
- 3 finished carousels
- HTML + PNG slides
- Platform captions
- Metadata JSON
- Queue entry in log
How it actually runs
Read the brain
Opens the memory file and last week's Tracker report. Identifies which hook patterns scored 7+/10 and which formats bombed.
Pick 3 topics
Priority order: Optimizer's recommendations → follow-ups on last week's winners → gaps in queue → evergreen expert topics.
Research in NotebookLM
For each topic, asks NotebookLM for Sam's real story, numbers, and take. Never invents a claim that isn't already there.
Write the copy
Hook under 12 words. Slide 2 standalone-hookable. Every slide one idea. CTA uses rotating comment keywords (STUDIO / AUDIT / COURSE).
Render + score
Builds the HTML carousel in Sam's design system. Runs the scorecard (80+ required). Anything lower gets rewritten or dropped.
Queue it
Writes PNG slides, uploads to CDN, drops into the queue folder with metadata. Logs the entry so the Scheduler can find it.
The rules I hard-coded in
- Write for experts and coaches. Never photographers.
- Every post needs CONTRAST (two worlds colliding, unexpected angle).
- Fact-check every number against real content. No inventing.
- Never claim client results I don't actually have.
- Reject any post scoring under 80 on the scorecard.
Files it touches
The Cutter
Scans raw footage in Google Drive. Picks the best moments. Removes dead air, grades color, burns captions + hook text. Saves finished vertical reels.
INPUT
- Raw .mov / .mp4 in Drive
- Modified within 14 days
- Current queue depth
- Top hook patterns (memory)
PROCESS
- Whisper transcription
- Filter expert-audience clips
- Write edit plan
- V2 word-gap cutter
- LUT color grade
- Dual-layer captions
OUTPUT
- Vertical 1080×1920 reel
- Dual-layer captions burned in
- Platform captions (IG/YT)
- Entry in reel_tracker.csv
- Dropped in "Ready To Post"
The 5-step edit pipeline
Transcribe
faster-whisper extracts word-level timestamps for the full clip.
Plan
Reads the transcript. Picks the viral moments. Writes an explicit keep/cut list.
Encode
V2 word-gap cutter removes dead air. LUT applied at 80%. 9:16 crop or horizontal-in-vertical for podcasts.
Overlay
Pillow renders hook PNG + auto-caption overlays. Dual-layer: bottom captions + center hook text.
Verify
Extracts a test frame. Confirms LUT, hook readability, no clipped words. Logs to tracker CSV.
+ 04
The Schedulers
Two separate agents. One handles carousels, one handles reels. Each runs every single morning, picks the next piece from its queue, and schedules it via my social scheduler's API.
EMPLOYEE #3
Carousel Scheduler
Daily 7:30am · 1 post / day
- Read the queue. Pick oldest unscheduled carousel.
- Calculate today's 10am CET slot in UTC.
- Check CDN URLs exist for every slide.
- POST to scheduler API with Instagram account.
- On Mon/Wed/Fri, also post to LinkedIn.
- Move files from queue/ → scheduled/.
- Update log. Write to Apple Notes.
EMPLOYEE #4
Reel Scheduler
Daily 6am · 3 posts / day × 4 platforms
- Check posting log for already-used videos.
- List unposted reels in "Ready To Post" folder.
- Pick 3 for today's 8am / 12pm / 6pm slots.
- Generate captions (expert-audience framing).
- POST to scheduler API across IG + TikTok + YT + FB.
- Append to POSTING-LOG.md with post IDs.
- Alert if queue empty or videos fail.
The safety rules both schedulers enforce
- Never post to Sam's personal Instagram (hardcoded block).
- Never post to any client account (hardcoded block).
- Never exceed 1 carousel per day per platform.
- Never exceed 3 reels per day.
- If token expired → log alert, skip, don't crash.
- If queue empty → log alert to Apple Notes, skip gracefully.
The Tracker
Pulls analytics for every post that went live this week. Scores each one. Writes a full performance report. Updates the memory file so next week's content is smarter.
INPUT
- Scheduler API (last 14 days)
- CAROUSEL-LOG.md
- POSTING-LOG.md
- Current memory file
PROCESS
- Split into carousels + reels
- Calculate engagement scores
- Apply format bonuses
- Rank hook patterns
- Identify top/bottom performers
- Cross-reference logs for context
OUTPUT
- Performance report (Apple Notes)
- Optimizer input note
- Updated memory file
- Moved posts → "Posted" archive
The scoring formula
Carousel Score (0–10)
Base: (likes + comments) normalized vs this week's top carousel × 10.
Bonus:
- Comparison format → +1.0
- Level-progression → +1.0
- Personal story → +0.5
- Tutorial → +0 (baseline)
Reel Score (0–10)
Base: (likes × 1 + comments × 3) normalized vs top reel × 10. Comments weighted higher because they're rarer and drive DM flows.
Bonus:
- Price contrast → +1.5
- Raw / single-word → +1.0
- Label-colon hook → +1.0
- Question hook → +1.0
The 4 questions the report must answer
- Which hook pattern is winning right now?
- Which format should we make more of?
- Which topic bombed and should we drop?
- What should next Monday's Creator prioritize?
The Optimizer
This is the one that closes the loop. It reads the Tracker's report, then edits the other agents' instructions. New hook priorities, updated content ratios, fresh topic recommendations. The team gets smarter every week. Automatically.
INPUT
- Tracker's weekly report
- Memory file (all weeks of data)
- Current Creator instructions
- Current Cutter instructions
PROCESS
- Count data points (≥3 weeks?)
- Rank hook patterns by score
- Calculate format averages
- Find cross-format winners
- Generate topic recommendations
OUTPUT
- Updated Creator skill file
- Updated Cutter skill file
- Updated memory file
- Content Recommendations note
- Optimizer Run summary
Decision logic
The hard guardrails
- Only updates content ratios, hook rankings, topic recs. Never touches design system, APIs, or safety rules.
- Needs minimum 3 weeks of data before changing ratios (prevents noise reactions).
- Documents every change in the skill_update_history log.
- Max any single format ratio: 0.60. Min: 0.05. Never lets one format dominate or disappear.
- Never reverts a previous optimization without 3+ weeks of contrary data.
Example of what the Optimizer actually says
"Last week your comparison posts averaged 8.2/10 and your tutorials averaged 4.1/10. So next Monday I'm telling the Creator: 60% comparisons, 20% tutorials, 20% level-progression. I'm also adding '$5K client vs $50K client' to the top of the hook rankings."
One shared memory. Every agent reads and writes to it.
This is what separates a team from 6 disconnected scripts. The memory file is the nervous system. Every agent leaves traces. Every agent can read the full history.
Weekly Reports
Every Friday's scores, hook patterns, top + bottom performers. Time-series of what worked.
All-Time Patterns
Running averages across every week. The multi-week signal, not the single-week noise.
Content Ratios
The current mix: 40% transformations, 30% tutorials, 20% contrarian, 10% personal. Adjusted by the Optimizer.
Hook Rankings
Which hook styles work best, split by carousel vs reel. Refreshed every Saturday.
Skill Update History
Every change the Optimizer made, with a timestamp and reasoning. Audit trail of the AI editing itself.
Cross-Format Insights
Topics that perform in both carousels AND reels. These are gold. Everything else is context-dependent.
One AI doing everything
will always be mediocre.
Six AIs, each with one job,
handing work to each other.
That's a team.
The magic isn't the AI. It's the handoff. It's the memory file they all share. It's the feedback loop that makes them self-correct.
That's how you turn AI from a tool into an employee.
I hated school.
I think certificates are fake.
I spent years staring at a clock in a classroom learning things I never used. I watched my friends go to university to spend 4 years and thousands of euros getting a piece of paper. I think it's a scam.
When I hire people, for myself or for my clients, I never ask for a certificate. I look at three things: the results they've already gotten, their character, and their charisma. That's it. No one has ever asked me for my high school diploma.
What school teaches
How to be graded.
What pays the bills
How to get paid.
My mission: teach people how to become valuable in the age of AI. Not in 4 years. Not with a degree. Right now. I'm building this whole system in public so you can copy it, break it, improve it, and use it to build your own income. Without asking permission from a school, a boss, or an algorithm.
That's why this page is free. If one person reads it, rebuilds a version of this for themselves, and gets paid for real work. I've done my job.
Build your own team in 6 steps.
You don't need my exact stack. You need the pattern. Here's the pattern.
List every task you hate repeating
Write down every weekly/daily content task. Posting. Editing. Captioning. Reporting. Researching topics. This is your org chart.
Give each task one "employee"
One agent = one job. Don't make a mega-agent that does everything. It will forget rules and drift. Split ruthlessly.
Write one clear job description
Every agent needs: inputs, process, outputs, guardrails. Be specific. "Post to Instagram" is not a job description.
Put them on a schedule
Cron them. Daily, weekly, whatever. The second you have to press "go" manually, the system is dead.
Build one shared memory file
A single JSON or markdown file every agent reads and writes. This is the nervous system. Without it, you have 6 amnesiacs.
Add a Tracker + Optimizer last
This is the most important step. Without the feedback loop, your team is just automation. With it, your team learns.
The best don't hide the recipe. I'll build you the restaurant.
Everything on this page runs my real business every day. You can copy the recipe piece by piece, or I can build the whole thing for you, in your voice, and run it.