#learninginpublic

The record started without permission

I did not begin by making a journal. I began by trying to get work done, which is how most of my systems have started before they developed opinions about my life. ChatGPT was where I put the question when I did not yet know what the question was, where I tried names for things, where I tested whether an idea had enough structure to survive daylight. It became a record because I kept returning to it with unfinished decisions.

That distinction matters to me, because a journal sounds intentional in a way this was not. I was not sitting down to preserve my thinking for later study. I was trying to make the next decision less slippery, and the side effect was a long trace of how I actually moved from one piece of work to another. The archive was not a shrine. It was a workbench with excellent memory and no concern for dust.

Only in hindsight did I recognise the shape of it. The chats were not just outputs, prompts, drafts, or notes; they were where I had been negotiating with myself in public enough for a machine to answer and private enough for me to keep going. A normal notebook waits for me to return to the page. This one answered back, occasionally with great confidence, which meant I had also invented a notebook that could be wrong at speed.

By the time I noticed the journal, I had already flattened too many ordinary turns into summaries that could travel cleanly

I had already been leaving

The ending was not sudden when I look at the behaviour closely. Long before I moved the work, I had been rehearsing the move in miniature. A chat would grow too large, continuity would begin to drift, and I would start another one, carrying context across by hand like someone moving house with one unreliable cardboard box. I called this continuation at the time, because calling it migration would have made the whole thing sound more planned than it was.

That repeated transfer changed the record before I admitted it. Every new thread required me to decide what mattered enough to bring forward, what could be compressed, and what could be left behind without breaking the work. I was not only preserving context. I was editing my own history into the smallest form that would let the next conversation function, which is a practical habit and a slightly severe way to treat evidence of a life.

The odd part is that I trusted the copied context more than the original sprawl. The old conversations had all the mess, but the hand-carried version had the working shape. I kept making little fences around the machine so it could sound useful without wandering off, then I acted surprised when the fences became the important part. This is one of my steadier talents: I set out to improve the room and eventually discover I have been measuring the door.

I kept confusing continuity with successful packing, as if a life stayed intact because the labels survived the box. The cost was quieter than I wanted to admit: I became too fluent at leaving context behind, and that fluency made each small loss look procedural.

The decision was less romantic

When I finally moved my working life from ChatGPT into Claude, the decision was not a dramatic preference about which tool had the nicer voice. That would be a cleaner story, and also a less accurate one. I chose Claude over Codex as the main workflow controller because the record already said the practical reason: stronger permissions, hooks, and workflow automation. I needed the place where I thought to be closer to the place where work could actually happen.

That changed the meaning of the old chats. In ChatGPT, I had been discussing, framing, testing, revising, and moving context from one container to another when the container became too heavy. In Claude, the work had firmer ground under it. It could sit closer to files, rules, and actions. I am deliberately not describing the mechanism, because the wiring is not the subject here and I have already spent enough of my life making cables look like destiny.

The important thing is that I did not move from one notebook to another notebook. I moved from the place that had been recording the work into the tool that watches me work. That is a different kind of relocation. A notebook observes after the fact, if I let it. This system observes as part of the doing, which makes the old record look less like a memory palace and more like a waiting room with good chairs.

The old failure finally closed

There was an older pattern hiding inside the decision. My first system was built to remove the copy-paste and it ended by choosing the copy-paste. That line has followed me because it is almost too neat, and because it is true in the least flattering way. I had tried to bridge two AIs, failed to make the bridge behave, and then built the workflow around the manual transfer I had wanted to avoid.

The migration to Claude did not solve that by building a better bridge. It solved it by making the bridge unnecessary. I stopped carrying the work across because I moved into the place where the work could be done. The old copy-paste died quietly, without a ceremony, which is exactly the kind of result my systems prefer. They do not announce success. They just remove the awkward bit and leave me with fewer excuses to make a diagram.

This is why the decision matters more than the tool comparison. A comparison would ask which interface was better, which answer was stronger, which model behaved more usefully. I was dealing with a simpler and more intrusive question: where does my thinking live when the place that records it is no longer the place where the work happens? Once I saw that, the migration stopped looking like a preference and started looking like an address change.

The record ended because I stopped exporting myself to be understood, and started building in a room where the export was no longer central. I had mistaken the missing entries for absence, when they were really evidence that the workflow had finally swallowed its own transcript. What stopped was not the work, but the separate act of carrying proof of the work back to a place built for asking.

The last entry changed the map

The ChatGPT record ends at 2026-06-19. That is the hard edge in the harvested timeline, the last dated entry in the place that had been carrying my thinking for more than a year. Nothing about that date says I stopped. The work did not disappear, the questions did not close, and the habits did not retire. The line ends because the activity that produced the line had moved somewhere else.

That is the part I had to let land without tidying it into a productivity lesson. The archive did not fail to keep up with me. I left it. I moved the working life out of the place that had been recording it and into the tool that was watching me work, and the old record went quiet because there was no longer enough thinking happening there to keep it alive.

The journal did not end because I stopped. It ended because I moved in.

The answer changed what I keep

So the answer to the opening question is not that the record ended because I abandoned it. It ended because it had finished being the room where decisions were made. A record is only worth keeping while it is close to the thinking it claims to hold, and once my thinking changed address, continuing the old record would have been a performance of continuity rather than continuity itself.

I still keep the old chats as evidence, but I no longer mistake them for the current map. They show how I got here, not where the work now lives. That difference matters because my working life has moved into a watched room of my own making, and the honest record is no longer the one that says the most. It is the one that stands closest to the work while the work is happening.

The checks passed, and that was the problem

I had trusted the green marks in the ordinary way, which is to say I had treated them as evidence that the thing was right, rather than evidence that the thing had performed the particular dance I had taught it to perform. That distinction sounds obvious after the damage has been found. It was less obvious while the system was doing what I asked, producing output, and allowing me to continue my day under the pleasant fiction that completed work and correct work are relatives.

The tests were not useless. They were obedient. They checked that the exclusion step ran, and that the pipeline completed after it ran, which meant they could certify a bad decision with excellent posture. A shorter list looked exactly like a cleaned list. A missing thing did not crash anything by being missing. It simply stopped being available to learn from, which is a tidy form of losing evidence.

This is one of the hazards of building systems that tidy up after themselves. When they succeed, they remove the trail. When they fail, they may also remove the trail, but with more confidence.

The cost was a kind of clean ignorance: I could repeat the run, compare the counts, and still not know what had vanished

I asked the smaller question

The useful part did not begin as an audit. I was not being thorough, noble, or even especially suspicious. I saw a count that looked too neat, too concentrated, too willing to behave, and I asked what had actually been removed. That is all. No grand method announced itself. No serious investigation stood up and cleared its throat. I had just learned, slowly and at some cost, that a number can be less a measurement than a door handle.

What made the question worth keeping was how little it required from me. I did not need a new framework for doubt, which was fortunate because I would probably have named it badly and then spent the afternoon maintaining the name. I needed to pause over the cheap question instead of walking past it. What are the missing things? Who decided they were missing? What was the decision actually based on?

I have learned to respect the questions that look too small to file properly.

The list was too tidy

Once I looked, the shape was wrong before the details were. The supposedly obsolete things were not scattered around the way real age scatters through a working system. They were clustered. They lived in one neighbourhood, as if obsolescence had developed a local council and kept minutes. That mattered because reality is usually messier than a pattern match, and when a failure is too tidy, I have learned to stop admiring the tidiness and ask what made it.

The rule I had wanted was plain enough: keep the system from learning dead material. The rule I had actually built was cruder. It looked for a warning word anywhere on a documentation page, then treated the thing on that page as if it had declared itself obsolete. That sounds almost reasonable if I say it quickly, which is where many of my worst decisions prefer to live.

A page can mention a retired setting while describing a current, ordinary, heavily used thing. A document can contain a warning without being the warning. The difference between mentions the word and is the thing was not a philosophical gap. It was where the system had been throwing away useful material while my tests sat there with clean hands.

The mistake was cheaper than judgement in the moment, because looking for a word was easier than asking what the page was actually saying. I saved myself the work of reading context, then paid for it by making the system forget material it was meant to learn.

The old rule had no teeth

The part that made the mistake mine was not the crude check by itself. I had already written the principle that should have stopped it: never drop data silently. I had the rule. I had the sentence. I had the satisfaction of having noticed the right danger, which is a surprisingly efficient substitute for preventing it. What I did not have was a check that forced the system to obey the sentence when the sentence became inconvenient.

That is how a rule becomes decoration. It sits in the project like a smoke alarm with no battery, adding moral architecture to the room while the actual machinery gets on with its private hobbies. I had built a fence around the behaviour in my head, then failed to put the fence where the behaviour lived. The result was not dramatic. It was worse than dramatic. It was administratively wrong.

In June 2026, the fix was not to become more suspicious of every result in some theatrical way. It was to make the rule enforceable at the source. A thing only counted as obsolete when it declared itself obsolete. Ambiguous mentions kept the item. If the system was unsure, it preserved evidence instead of deleting it and asking me to be grateful for the tidier floor.

A wrongly kept item can still be seen. A wrongly dropped one becomes a silence with consequences.

The question caught the number

The question in my log was: "What are these dropped (deprecated) nodes and how were they decided?" When I followed it, 196 were on the list and only 2 belonged there. The other 194 were wrongly flagged.

That was the point where the green checks stopped meaning what I had been letting them mean. They had not lied. They had answered a smaller question than the one I needed answered. They could tell me that exclusion happened, and that the pipeline survived it, but they could not tell me whether exclusion had become a quiet little machine for deleting the present because a page mentioned the past.

The repair did two things at once. It narrowed the judgement so the system only acted on a self-declaration, and it changed the default when the evidence was unclear. Keep first. Remove only when the claim is specific enough to deserve removal. I also added a guard for recurrence, because I had already proved I was capable of writing the right principle and then leaving it to supervise nothing.

I kept the questions

I would prefer the story in which I found the problem through disciplined engineering pressure. It would be cleaner. It would make me sound like the sort of person who simply aims rigor at a system until the system confesses. What actually happened is less flattering and more useful: I was curious on a day when I was not formally looking, and the idle question found a defect that the formal checks had approved.

So I changed the habit. I stopped treating doubt as a mood I had to either indulge or overcome, and started treating it as an instrument I could keep on the bench. Not every question deserves a hunt. Some are noise, some are procrastination wearing a clean shirt, and some are just my brain trying to avoid the next hard thing. But the cheap questions that keep finding expensive mistakes have earned a place in the work.

That is how I catch a mistake that does not look like one: I keep the question long enough for it to become testable, and I make the system answer the question I actually meant to ask.

I carried forward the old backend habit after all: make the invisible decision visible enough that it can be argued with. What changed was not the amount of doubt, but its route. There was less of it circling in my head, and more of it pressed against the machinery, where Houdini could expose whether the decision actually held.

I kept losing the work

There is a particular kind of failure that does not look dramatic from outside. I am not blocked at the start, when the page is blank and everything is still possible. I am blocked after I have already done enough work to make the next choice matter, when the screen has evidence on it and I can no longer hide behind preparation.

That was the place I kept circling. I would be inside creative software with something half-made in front of me, and the problem would refuse to identify itself. Sometimes I thought I needed a technical fix, sometimes I thought I needed a visual correction, and sometimes I was only watching my own judgement wobble because the thing had reached the point where the next move could improve it or make it worse.

Leaving the work was always available, which is why the problem took me longer to recognise. I could go and find a tutorial, ask a model, search through notes, or turn the whole thing into a separate research task, but the cost was that the actual moment had already passed. By the time I came back, I was no longer solving the same problem. I was solving a description of it.

I kept turning live decisions into homework, and every time I came back the work felt slightly less mine, as if the problem had been translated once too often before I touched it again. The mistake was counting the search as progress, when really it had only protected me from making the next visible mark.

The idea waited without a system

I did not begin with a product plan. I began with a pull that kept taking the same shape, even when I did not have the tools, habits, or craft to make anything from it. I wanted something beside me while I worked, something that could see enough of the work to make its advice belong to the moment instead of arriving afterwards as general advice.

At the time, that was not a build. It was only a description. I had no working system, no useful architecture, and no clear path from the sentence in my head to a thing I could actually use. So the idea did what a lot of ideas do when they arrive before the hands are ready for them. It stayed as a shape and let other work take over.

That other work looked unrelated from the outside. I worked on the portfolio, pushed deeper into procedural craft, and started the MA, and none of that had the neatness of a system slowly assembling itself in public. It was messier than that. I was learning how to make, how to judge, how to keep a creative problem alive long enough to understand it, and how to stop treating technical structure and visual work as separate rooms in my head.

The idea did not disappear because I had rejected it. It disappeared because I had no place to put it yet.

I built the wrong answer first

The first real response I built to that pull was AI Central, and I wrote about that failure in The First System I Built Was Wrong: The Guardrail Tax. It was a coordination hub, which means it treated the problem as if the work needed to be managed from above, organised, routed, and contained before I was allowed to return to the actual making.

That diagnosis was close enough to be dangerous. I did need structure, and I did need a way to stop my tools from scattering across the desk of my attention, but the deeper problem was not that my creative work lacked a command centre. The deeper problem was that I kept reaching the point of decision alone, with the work open in front of me and guidance living somewhere else.

AI Central retreated within two weeks. That matters more to me than the fact that it failed, because the retreat showed the shape of the mistake. I had built something that could oversee the work, when what I had been asking for all along was something that could stay beside it.

The mistake was useful because it made the wrong scale visible: I had solved for organisation when the break was happening inside judgement. Once the hub failed, I could stop asking how to gather every tool and start asking what needed to remain close to the screen.

The right answer arrived beside the work

What came next was True Tutor. It became an always-on-top window that watches work in creative software, checks it at checkpoints, and answers doubts while I am still inside the making.

I am keeping that description short because the mechanism is not the point here. The point is that the system finally matched the old pull. It did not ask me to leave the work, translate the work into a separate problem, and return with advice that had lost contact with the actual screen. It stayed close enough that the question could remain practical.

That changed the nature of the help I was trying to build. A tutor beside the work is not just a smaller hub, and it is not a friendlier interface on top of the same idea. It starts from a different belief about where creative judgement breaks. I was not trying to make a smarter archive of what I had done, or a better command system for what I should do next. I was trying to make the moment of uncertainty less isolated.

I found my own words waiting

Only afterwards did I go back through my own old conversations and find the sentence. That is the part I keep returning to, because I did not find a clever plan that I had faithfully executed. I found a version of myself describing the thing before I could build it, then losing sight of it while I did the work that made it buildable.

The concept note was dated 2025-05-07, and the first attempt began on the same calendar date a year later. The coincidence is what my records show; I am not claiming fate.

The line was sitting there plainly: "Explored the possibility of AI systems that can analyze screen activity, provide workflow suggestions, and automate tasks across different software."

There was no system attached to it. No tools listed. Nothing built. Just the shape of the desire, written early enough that I could no longer pretend the tutor was only the result of a recent technical experiment. I had already named the thing I wanted. I had just not become the person who could make it useful.

Reading that old sentence back, I could see that the answer had never really been more information, but a way for the work to stay present while guidance arrived, close enough to matter and specific enough to change what I did next. I had not invented the need recently, I had only built enough around it for the need to stop sounding abstract, and once I understood that, the opening question changed shape. I was no longer asking where to find help, but where help had to stand.

The question finally closed

So the question I opened with has a cleaner answer now. How do I get guidance at the moment of making, without leaving the work to go find it? I had to stop treating that as a search problem, then stop treating it as a management problem, before I could recognise it as a presence problem.

The idea came before the system because the idea was not waiting for the right date. It was waiting for enough craft, enough failed structure, and enough time spent getting stuck in the same place to show me what kind of help I was actually asking for. AI Central was the wrong answer to the right pull. True Tutor was the same pull finally put in the right position.

What it was supposed to be

I called it AI Central. The idea was a hub sitting above the systems I already had, connecting the creative work, the publishing side and the pile of research. It would coordinate, cross-check, route a task to the right place, and remember what had been decided and why. It was never meant to do the work. It was meant to know where the work should go.

On day one I listed the nine areas it would connect: research, code execution, tooling, publishing, documentation, my 3D workflows, portfolio, social, and a feedback loop that would learn my preferences. Reading that back, the ambition is obvious and slightly embarrassing. At the time it read as organised. Two weeks in I quietly downgraded it in my own notes to a setup hub rather than the permanent controller, and I treated that as scoping rather than as a warning.

It worked, and it did almost nothing

The folder built to hold the research, the entire reason the thing existed, contains nine files. Every one is a README, an index or a guide. There is no research in the research folder, only a description of how research would be filed if there were any. The folder for completed tasks holds a README explaining what a completed task would look like. It never received one.

The bridge was the one object the system existed to produce: the thing that carries a job from one place to another. Its index is still on disk. Under "Active Bridges" it says: "None currently." One was ever made, on day two, annotated as a test not to be run for real. That was the high-water mark.

Meanwhile 87 percent of the files sit in the plumbing, including 136 planning documents for a web interface. One hundred and thirty-six, for a hub that would coordinate work which was not happening.

The rule I wrote before I wrote any code

Here is the twist, and it took me a year to see it. The safety rule was not the lesson. It was the premise. It is sitting in the kickoff notes, before a single folder existed: a review-and-approve production assistant, not an automatic bot; never publish, send, submit, delete or overwrite anything final without my say-so.

I still believe every word of that. I have not changed it once since. It is also what killed the system, for a reason that is almost stupid. I wrote a rule about publishing, and then I applied it to everything.

Every internal step inherited it. Writing a note to itself needed approval. Reading a file and drafting a summary needed approval. Starting its own local process needed approval. A system that never published anything in its life was carrying publishing-grade caution on every move it made, and every move cost a round trip through me.

That is the tax. Not one blocker you can see and remove, but a small toll on every single action, paid in my attention, until the cheapest thing the system could do was write a document about what it would have done. So that is what it did. One hundred and thirty-six times.

I noticed, and I lost anyway

The record shows I was not oblivious. Five separate times I went in and patched it: fast-track the approvals, prefer doing over reporting, stop the report-only planning loops, replace the source files that kept dragging it back toward planning. At one point I had to write this down as a plain-English instruction: when I ask you to research or check something, do not automatically turn that into a report about researching it.

It did not take. I had built something that out-cautioned its author, and then I argued with it five times and lost. The disease was not in any of the places I was patching. It was in the premise, and the premise was mine.

I kept telling myself the next patch would be the one, even though I could feel the problem was structural, and I kept choosing the smaller repair anyway because symptoms are the kind of thing you can still fix in an evening.

How it ended

The last substantive thing in that folder is a fix to its own web interface, so that I could look at it. The fix worked. The page loaded. Then it found that a leftover process from an earlier run still held the port, so it stopped and wrote down its options. One of them was to explicitly approve stopping that exact process.

So the final act of my first AI system was asking permission to stop a copy of itself, so that it could look at itself. I never gave it. Nothing dramatic happened. There is no decision anywhere saying I shut it down. It just went quiet, and the creative system it was built to coordinate, which had existed a day before AI Central did and never needed a hub, carried on without it.

One more detail I find hard to look at. The whole point of that web interface was to stop me hand-carrying text between two AIs. Its final recorded decision was to select manual paste as the safest option, with me as the executor. It was built to remove the copy-paste and it ended by choosing the copy-paste. Roughly 350 steps to arrive back at the start.

What I actually got wrong

For a long time I would have told you the lesson was to decide what the AI does and what I approve. I said it for months. It is wrong. It is the mistake wearing the costume of the lesson, because that boundary is what I started with, and it is what did the damage.

The question is not whether the machine acts. It is what the action can break. Drafting a paragraph in a file I own and publishing that paragraph to the internet are not the same kind of act, and treating them the same is not caution. It is a failure to think, dressed up as caution. Guardrails that fire on everything are not safety. They are noise, and noise gets ignored or routed around, which is worse than having no rule at all.

I only admitted that a few days ago. I went back to the dead system and had it audited against what I know now, because I had a suspicion I did not like: that it had been over-protective from the start and I had been calling that responsibility. The audit agreed with the suspicion. The fix is not to remove the guardrails. It is to make them tell the difference.

What I missed was that a rule I had never broken was almost invisible to me, because nothing ever went wrong loudly enough to force my attention onto it. I think I needed the second system to work before I could afford to admit why the first one did not, which is uncomfortable, but also useful, because it means the failure was not just in the code.

The rule I have not broken since

Gate on what it can break, not on who is doing it.

Reading, searching, inspecting, drafting something locally where I can see it and undo it: that should never stop and ask. Changing something that runs, or that other things depend on: worth a question. Deleting, sending, spending, publishing, anything touching credentials: that stops every time, no exceptions, and I would rather it stopped too often than once too few.

The day-one rule survived intact. Not one word of it changed. What changed is that it stopped applying to everything and started applying to the short list of things that are actually irreversible. The system I use now is freer than AI Central ever was, and more locked down where it counts. Those are not in tension. I could not see that when the only dial I knew about was how much I trusted it.

I have had to admit that the failure was worth more than the system would have been if it had worked, because a working version would probably have let me keep believing the wrong thing with more confidence.

I keep the old folder. Its rulebook still sits in my current project, still describing a system that has not run since June: a coordination hub that never coordinated anything, explaining its constitution to nobody. I have not deleted it. Deleting is on the list that stops and asks, and I have not answered.

It worked by disappearing

Backend work taught me to care about things nobody looks at directly. A report arrives, an endpoint responds, a pipeline runs, a monitor stays quiet, and if the work is good enough the person using it never thinks about the structure underneath. I was good at that. I liked the precision of it, and I liked the way a small decision in one place could either make the rest of the system easier to reason about or quietly punish everyone later.

That is a strange apprenticeship for someone who later decides he wants to make visible things. I had become used to a kind of craft where success removes the maker from the evidence, where the best compliment is often silence, and I carried that into a field where a blank page does not congratulate you for uptime. The server stays quiet because it works. The sketch stays quiet because there is nothing on it yet. These are different silences, and only one of them can be billed.

The problem was not that the work was meaningless. It was that I wanted to make things a person could actually look at, and that want did not behave like a career preference. Sketching and art had been there before the job, then gone quiet inside the Monday-to-Friday, 9-to-5 routine, until the pressure became less like a hobby asking for time and more like a simple refusal to stay separate.

I learned to accept outcomes I could explain but not point to, and that made my own taste too easy to postpone. I let usefulness stand in for authorship for longer than I should have, because useful work rarely asks whose hand is missing.

I became a beginner badly

Being a beginner is different when you have recently been paid to know things. I could not hide inside general curiosity, because I already knew what competence looked like from the inside, and I knew the difference between confusion that is useful and confusion that is just lack of practice. Alongside a summer as a packaging-design intern, I was teaching myself the visual stack from scratch: After Effects, Illustrator, Photoshop, Blender, and all the ordinary humiliations that come with trying to make the hand obey the eye.

That year did not behave like a clean switch from one life to another. I was not walking out of code and into art with a tidy little bag of transferable skills; I was bringing habits from one craft into another before I had the language to say what I was doing. Every tutorial became less about copying an outcome and more about finding the reusable handle underneath it, the part I could change later without rebuilding the whole thing. This made learning slower, which was not the advertised benefit of curiosity.

I kept distrusting the finished image.

What I trusted was the structure that had produced it. If I copied a look and could not alter it, I had learned a pose, not a method, and I had already spent enough time around systems to know the difference between something that works once and something that can survive being touched again. I did not have the vocabulary yet, so I treated every small visual problem like a backend problem wearing brighter clothes.

I had to lose the habit of being efficient, because the shortcut usually kept the weakness I needed to draw through, just hidden in a cleaner place. More than once I mistook software progress for artistic progress, then had to admit that a tidier file could still be carrying the same bad decision.

The graph changed the question

The clearest turn happened inside node graphs. In Geometry Nodes, and later in ComfyUI while building a character sheet, I noticed that I was not really making the picture in the way I had imagined artists made pictures. I was making the thing that made the picture, then changing the behaviour of that thing until it started giving me something I could argue with. The graph was the brush. A parameter change was the stroke.

This should have warned me. I had left invisible systems and immediately started making small visible systems, except now they produced faces and terrain instead of reports and endpoints. The surface had changed enough to fool me for a while, but the working habit had not. I would make a version, notice what rule inside it had misbehaved, change that rule, make another version, and then take the machine more seriously than the output. It is a very direct route to art, provided one defines direct as building a mechanism, questioning the mechanism, and only then looking at the picture.

The MA and Houdini sharpened that turn because they gave me a place where the rules were not only hidden inside someone else's tool. I could decide what a thing should be allowed to do and encode that decision, which is a different pleasure from making the asset itself. The point was not control for its own sake. The point was that behaviour could become material, and I recognised that material more quickly than I recognised the image.

The portfolio argued back

The portfolio became another clue. I could have made a clean gallery and placed the work in it, which is what the word portfolio had been quietly requesting, but I built an interactive 3D site in React and Three.js, learning both while making it, because I wanted the thing itself to argue that the code and the art were not two separate people. I set out to make a container for the work and produced a system that behaved like one, because apparently even my evidence needs architecture.

That mattered because the portfolio was not only display. It was a claim about authorship. I was trying to make the visitor move through the same contradiction I had been living inside: the visual work was not an escape from software, and the software was not merely scaffolding for the visual work. I had built a place where the argument sat in the behaviour of the page, not in a paragraph explaining it, which was probably the first time the problem used the right medium to answer back.

The rules found their names

The AI thread began more ordinarily than it now sounds. I used ChatGPT as a notebook for modules, then as a research assistant, then as a way to turn tutorial transcripts into something buildable. That was useful enough, but usefulness was not the part that held my attention. I kept adding fences to a machine that could sound certain, writing down where its authority ended, deciding what it was allowed to conclude and what it had to admit it could not see.

Then the old habit reappeared. I was not only using a tool; I was designing the workflow around the tool, then the system around the workflow, then the boundary around the system. The first serious attempt was too ambitious and collapsed under its own safety rails, which is a very tidy failure for something built to prevent failure. What survived was simpler and stricter: the machine drafts, gathers, checks and organises; I decide.

The medium changed completely; the instinct did not move at all. I kept arriving, alone and from a different road, at ideas that already had names. One rule was "never guess, and mark what you cannot see". Another was "nothing acts without my approval". I am not claiming I invented either. I am saying that I wrote them because I needed them, then later discovered they belonged to a larger conversation.

I had carried a bias for systems that could be inspected, changed and argued with, even when the output was no longer software. I had carried the part of software that asks what must be true before anything is allowed to run. I had carried a refusal to let tools become authorities, and although that could sound like control, the work made it visible to me as judgement.

So I am writing it down

That answers the opening question as directly as I can without making the answer neater than the life that produced it. I did not leave one self behind and assemble another from better software and better images. I carried a way of thinking across surfaces, and each surface made one part of it visible enough for me to recognise. Backend work taught me to care about hidden structure. Art made the structure answer to sight. AI forced me to write the boundaries down.

This blog is where I want to put that habit under pressure. The art lives elsewhere. This is the workshop, the place for short posts about questions and what they caught, decisions and what they cost, failures in full, and the small rules I keep finding after I have already started obeying them. Most of this has been me alone in a room, testing ideas against my own judgement and hoping I am not just getting better at convincing myself. That is a poor review process. I want to be argued with.