Not because I had a framework in mind. Because I wanted to see how far I could push it. What could actually be automated. What couldn’t. Where it made me faster and where it just made me wrong more efficiently. Bjørn started as a personal experiment. A way to learn new skills, to test, to find the edges. It is still a work in progress. This is the story so far.
“When AI does the work, what do you do?”
That question is not rhetorical. It is the one I kept coming back to as I built this. And the honest answer, at least at first, was: not much different. Slightly less admin. Same decisions. Same bottlenecks. Just faster. The three stages below are how I think about the progression now — not as a framework I designed, but as a pattern I kept running into until I gave it a name.
You give it tasks. It does them. You are still running the queue, deciding what matters, chasing the same priorities. The AI is faster and more articulate than a search engine but it is doing exactly what you tell it. I was here for longer than I expected. Automated execution feels like progress. It is not the same thing as automated thinking.
This is the stage I am still navigating. The AI starts holding context across time. It knows your goals, your constraints, what you actually care about. It surfaces things you did not ask for. It pushes back when you are about to make a decision that contradicts last week’s thinking. The shift is subtle at first. You stop managing the queue. You start setting the direction. That sounds simple. It requires rebuilding how you work almost from scratch.
You say what needs to be true in 90 days. The system works backward. Research, synthesis, drafting, scheduling, follow-up — delegated, sequenced, and quality-checked without you in every step. Your job becomes the decision at the start and the judgment call at the end. Not the work in the middle. I have had glimpses of this. Enough to know it is real. Enough to know I am not there yet.
Most people jump straight to Stage 3. That is where the expensive lessons live.
The goal was never efficiency. It was balance. Something that could hold the whole picture. Work, health, thinking, relationships. Without me carrying all of it in my head at once. When I first mapped it out I had three dimensions: Mind, Body, Soul. It felt complete. Then I noticed something uncomfortable. Every component was getting better at agreeing with me. Faster research confirming my existing thesis. Sharper writing reinforcing the same arguments. Smart, capable, and entirely incapable of telling me I was wrong. The fourth dimension is the one I almost missed. It is probably the most important one. You may already know the feeling. It starts with your model saying “you are absolutely right...” and never really stops.
The part that thinks. It holds a 31-day calendar view, flags conflicts before I hit them, keeps the weekly priorities honest and runs my ideation pipeline end-to-end. It monitors market signals and tracks the professional network. The goal is simple. Nothing strategically important should fall through the cracks because I was too busy answering emails. That used to happen more than I would like to admit.
The part that keeps running when I am not looking. Fitness tracking, daily session logs, encrypted backups every Sunday, a 60-minute heartbeat that checks in without being asked. It also covers the physical side. Sleep, recovery, energy. The Mind produces better output when the Body is not quietly failing. That connection took me an embarrassingly long time to take seriously.
The part that knows the difference between a work problem and a family dinner that cannot move. It holds personal context. Who matters, what is coming up, when to push and when to leave it. It also monitors my communication patterns and leadership tendencies over time, which means it occasionally tells me things I would rather not hear.
Here is the uncomfortable truth about a well-trained personal AI. It gets very good at telling you what you want to hear. Perspective is the counter. A Red Team Protocol that argues against my current thinking. A Serendipity Engine that surfaces ideas from completely unrelated fields on a schedule. A shortlist of external voices chosen specifically because they push back. The goal is not balance. It is productive friction.
The most important thing I learned is that autonomy has to be earned. You do not start with orchestration. You start with supervision. Every output checked. Every decision reviewed. It feels slow because it is slow. That is the point.
Lobster-Gate was not a technology failure. It was an autonomy failure. I gave the system control it had not earned. Too much, too fast, with no checkpoints.
Every output reviewed. Nothing sent or published without a human eye on it. Feels slow. Supposed to.
Day-to-day runs. I review output and exceptions. The system has a track record in specific areas. I trust it within those limits.
The system shapes the agenda. Surfaces gaps before I notice them. Challenges assumptions before I act on them.
I set the outcome. The system figures out the path. Only works if Levels 1 and 2 were built properly.
“I spent over a hundred dollars learning that gates are not bureaucracy. They are the thing that makes autonomy sustainable.”
One model trying to do everything is like hiring a single contractor to build a house, wire the electrics, and design the interior. The model stack below reflects two years of figuring out which cognitive jobs actually need specialisation — and which ones you can hand to the fast, cheap generalist without noticing the difference.
| Alias | Model | Primary Use |
|---|---|---|
| Light | Gemini 2.5 Flash | Daily chat, heartbeats, admin |
| Writer | Claude Sonnet 4.6 | Strategy, deep reasoning, high-stakes writing |
| Researcher | Gemini 3.1 Pro | Massive context, long-document research |
| Solicitor | Claude Opus 4.7 | Legal review, sensitive protocols |
| Strategist | ChatGPT 5.5 | Clinical synthesis, logical anchoring |
| Scout | Grok 4.2 | Real-time search, edge-case signals |
| Coder | Mistral Large | JSON, scripts, technical troubleshooting |
| Artist | Gemini Flash Image | Image analysis and generation |
| Intern | OpenRouter Free | Simple tasks, high-volume boilerplate |
| Tool | Purpose |
|---|---|
| Tavily | Quick factual queries, current news, verification |
| Exa | Analyst reports, expert opinions, niche research |
| Firecrawl | Deep analysis of specific URLs and content extraction |
| Brandfetch | Clean SVG/PNG logos for microsites and decks |
| Telegram | Primary real-time communication channel |
| iCal Feed | Calendar integration |
| Image Gen | Diagrams, concept art, and visual aids |
| Himalaya | Email reading and prioritisation |
| memory-core | Daily logs, long-term distillation, background synthesis |
This is the actual log. Some of it is progress. Some of it is expensive lessons. I have kept both because the failures are more instructive than the wins, and because anyone building something like this will recognise the pattern.
Comes online. Calendar, files, messages. Does what it is told. Supervised on everything. Useful immediately. Which made me overconfident almost immediately.
First time I said out loud what I actually wanted from this. Not a faster inbox. A system that could hold the strategic picture so I did not have to carry it all in my head. Built the intelligence repository. Introduced weekly priorities. Felt ambitious at the time.
First time the system had a shape rather than just a list of features. Calendar automation went live. Memory protocol introduced so it survived session restarts. This is the version I could actually explain to someone.
Built the nine-alias stack. Gemini Flash for daily operations. Claude Sonnet for high-stakes writing. Gemini Pro for deep research. Cost dropped. Quality on the things that matter went up.
Tried to automate the entire content pipeline in one go. YAML pipelines, chained model calls, zero checkpoints. Spent over a hundred dollars. Got hallucinations, generic output, and a debugging session longer than just doing it myself. Creative work resists rigid automation. That lesson cost me a weekend.
Added Perspective after realising the three-dimension system was getting very good at agreeing with me. First time the system had a philosophy rather than just a feature list. Still building.
