working notes / founder journey
This took 125 days. It also took four years.
I thought I was building an AI product. Eventually I realized I had been building pieces of it for years.
The four years were the methods, the failures, the client problems, the personal workarounds, the frameworks and the ways of seeing. The 125 days were when enough of the graph became visible at once to start engineering the thing underneath it.
The structure is still moving.
01 / Before Yu had a name
I kept ending up where the context transfer failed.
Building other people's software
Finding the states nobody knew they had to specify.
Tech Translator
Getting the whole idea out before a team built the wrong thing correctly.
Grant writing
Translating a business into the representation a funding system could understand.
Brand + strategy work
Turning someone's internal model into something another human or AI could actually use.
Learning + visualization
Making invisible systems visible enough to move through.
How do you make the important context travel?
People described the product they could picture. What decided whether a build succeeded was everything around it that never got said out loud. The work was pulling those parts into the open before a team spent three months building the wrong thing correctly.
- payment / tier states
- negative states and failure paths
- admin / control surface
- edge cases and assumptions between the words
You didn't say that, but you also don't know what to say.
Colors underneath white. Two outputs can look identical and be composed completely differently underneath. I have used that image for years. It is a way of seeing I carried in, not the moment Yu started.
02 / Then AI removed the friction
Speed exposed the problem.
AI became an extension of how I think. The pauses I used to spend searching, waiting or handing something over mostly disappeared, and work could move at the speed I was actually working at.
Then the paradox. Removing the friction did not remove the load. By May 2026 I was describing AI-assisted burnout as my own lived experience, not as a theory about other people.
At the same time I was carrying context by hand. Re-explaining how I work, re-establishing constraints, keeping several conversations in sync, rebuilding the same picture again and again.
I was already the manual version of the thing I was trying to build.
03 / Yu does not have one lightning-bolt origin
The roots do not collapse into one moment.
Introspection
Analyzing my own thought patterns.
Self-knowledge premise
You can't use these tools to help you if you don't know anything about yourself.
Assessment experience
Seeing how useful self-knowledge could change the way I worked.
Manual AI pain
The representation needed to persist and travel instead of being rebuilt every chat.
Earlier AI + neurodivergence work
Systems should adapt to different people rather than pull everyone toward one default.
These are all real parts of the record. Their exact relationship is still open.
04 / The graph starts moving
Ideas did not stay in the room where I found them.
The same week can contain very different rooms. Something surfaced in one room and turned up somewhere else entirely, sometimes as a product decision, sometimes as nothing at all.
input, interpretation, transfer, refinement, artifact or no change.
A Yu browser-extension mechanism came up. The reaction was that you could almost do the same thing somewhere else. It moved into a separate client product as a purchase intercept, and the response was positive.
Same day. One room held about twenty minutes of Yu. Another two-hour working session held none of it. Being immersed in Yu was not enough to make it enter every room.
A wide framing was often what got the meeting. Once I was in the room, the conversation could move across several live problems at once, which is part of why ideas crossed between them. That is a property of how the rooms were opened, not evidence that everything spreads.
05 / How I actually work
The method showed up before I had names for it.
Graph, not spreadsheet
Many partially connected nodes held at once. A local answer can be correct and still deform the larger shape.
Up and down the chain
Specific expression, underlying mechanism, sibling expressions, earlier and later evidence, consequence.
bidirectional thematic tracing
React to something
Externalize a candidate, notice the violation, reveal the hidden constraint, refine.
Rumble to Render
Model formation and ambiguity reduction are one mode of work. Execution is a different one.
State access
When the graph is loaded, momentum is intact and executive-function friction is low, I can traverse a huge amount of context quickly. An interruption can cost more than the minutes it takes, because the model has to be rebuilt.
founder-reported operating experience
Build the structure while building the thing
I build enough structure to hold what I know now. Then I keep moving until the structure is wrong. Then I rebuild it.
another way to view the same graph
06 / 125 days of moving the bottleneck
The roadmap was the thing currently stopping the next thing.
Can cognitive translation work at all?
Can it become repeatable?
Can people understand what Yu is?
Can they actually get through it?
Why are they getting stuck?
Do they understand enough about AI to perceive the value?
How do we teach without turning Yu into a course?
How do tiny interactions accumulate into something worth returning to?
How do we preserve evidence, corrections and ownership underneath it?
The structure kept changing because each answer exposed the next constraint.
07 / The things I was wrong about
The record is more useful when it can disagree with me.
A meeting felt enormous. Nothing traceable came from it.
I distrusted a program rule. I was wrong, and later said so.
Useful advice arrived. There is no evidence I processed it.
I said I took something out. The code showed it had been gated or renamed.
A push-back instruction worked. 24 days later it had become a reflex that fired on good ideas too.
An idea sounded promising. No downstream artifact was found.
Excitement is not consequence. Speech is not shipped state. Silence is not rejection.
08 / What Yu finally became
A representation of the person, not another place the person has to live.
A user-owned representation of how intelligence should work with a specific human.
Evidence and observations stay distinguishable from derived findings.
Corrections and provenance survive.
Instructions can be conditional, not one size fits all.
The representation can change as the person changes.
It is portable by design.
Specialist systems can contribute scoped evidence without owning the person.
Your Yu belongs to you.
The specialist app owns the intervention. Yu owns the continuity of the person.
09 / Still building
A living model should leave its questions visible.
What we know
- Context has to be inspectable and correctable.
- One flat instruction can misfire.
- People need enough of a mental model of AI to exercise agency.
What we are testing
- Whether a Yu-derived interaction policy reduces correction burden on unseen work.
- Whether it transfers across models.
- Whether the correction loop improves the representation over time.
What changed this week
- Source reconstruction is changing how this journey itself is understood.
- The evidence model is becoming more graph shaped.
- External events and badges are being separated from higher-order interpretation.
10 / Closing
For years I kept ending up downstream of context that failed to travel. Yu is what happened when I realized the missing context was not only about the product, the business or the task. It was the person.
The Founder Journey is a view of the graph. It is not the graph.