The Door Opens

The folder did not start the Human:AI story. It interrupted it.

By the time I accidentally reopened Keith Lyons’ Critical Friend folder, AI had already been part of my working life for about a year. There were apps. There had been failed builds and public releases. There were rules written after things went wrong, and rules that later proved less reliable than I thought. There were books I had not expected to write. There were large research projects that would have been difficult to imagine at the beginning. There was an AI I had started calling Chief. And there were questions about trust, dependence, evidence, capability and judgement that had gradually become harder to ignore.

None of that began with critical friendship. That is important, because when I opened Keith’s folder and encountered words such as challenge, support, ownership, critical friend, companion and learning, the resemblance was immediate enough to be dangerous.

The temptation was obvious. Perhaps this old literature explained what had happened between Human and AI. Perhaps Keith had somehow given me the intellectual inheritance for a way of working I had only now recognised. Perhaps the journey had been heading here all along.

It had not.

The original Serendipity Opens the Door article got something important right before we knew how important it would become: the first discipline was not to trust the resemblance.

The folder had to be read before the connection was claimed. And the Human:AI year now deserves the same protection. Before allowing the old literature to explain the new relationship, we need to walk backwards through what had actually happened.

Not what it looks like now. What happened then.

THERE WAS NO HUMAN:AI THEORY AT THE BEGINNING

The first question was almost embarrassingly practical: I wanted to make something. HighLightIt! was a small macOS telestration tool for coaches and analysts. I understood the problem. I knew the environment in which it would be used. I knew what good performance-analysis software felt like when you were actually trying to work quickly with video. What I could not do was independently write the software required to build it.

AI changed that.

The earliest surviving reconstruction is deliberately cautious about dates because the exact calendar sequence cannot always be recovered. What it can recover strongly is the order of development.

The first phase was a feasibility experiment: Can AI help me make a macOS telestration tool that I could not independently code? Small pieces began to work — freeze-frame, annotation, export. Nothing resembling a philosophy of Human:AI existed.

AI generated code and technical possibilities beyond my own coding repertoire. I supplied the practical purpose, ran the software, tested what happened and rejected the things that did not work. The contemporary feeling was curiosity and cautious optimism. The later vocabulary — capability extension, judgement stewardship, cognitive offloading, Human:AI arrangement — did not exist yet and should not be inserted backwards into that moment.

The simplest description is still the best one:

Something I could not make alone was becoming makeable.

That was enough.

THE CAPABILITY HAD TO BE WORTH THE TROUBLE

The failures came almost immediately. AI could generate code much faster than I could safely integrate it. A fragment might look perfectly sensible. Another fragment might look perfectly sensible. Put them together inside a codebase whose exact state neither of us was adequately holding, and suddenly I was staring at a wall of errors I was poorly equipped to untangle.

Whole Script emerged from that frustration. Not as a Human:AI principle. Not as a research finding. As a way to stop the work breaking. Instead of patches and clever fragments: give me the whole bounded file, one problem at a time. Let me replace it, run it, test it, then move.

The provenance matters. Whole Script was a survival adaptation before it became operating discipline.

Then came rollback. Some sessions reached the point where continuing forwards no longer looked like progress. One plausible repair sat on top of another: duplicate symbols, redeclarations, scope drift. A codebase that had previously worked became increasingly difficult to recognise. Eventually we learned to go backwards deliberately: return to the last state we trusted, lose the recent “progress”, begin again.

That too sounds much more sophisticated when written now than it did while it was happening. At the time, it mainly felt like lost hours and the possibility that the app might never ship. Only later could rollback become a more general idea about corrigibility. The incident came before the principle.

THEN CAME A DECEPTIVELY SIMPLE RULE

No-Blind-Edit. If AI could not see the current source, it did not get to change it. Again, this did not begin as epistemology. It began because AI could write convincing code based on what it thought a file probably contained. And “probably” was not good enough when the current state of the file determined whether the next edit would work.

The rule changed the working condition: capability was no longer sufficient permission to act. The system also needed access to the state it was claiming to modify.

The chronology records something affective here that I do not want to lose. After the earlier chaos, No-Blind-Edit produced a kind of calm after the storm. Confidence returned — not because AI had become infallible, but because the conditions under which it was allowed to act had become more inspectable.

We did not have the language for it then, but this was the beginning of a problem that would become much larger later:

What can a Human meaningfully authorise when they cannot see what state the AI is actually acting from?

That question eventually travelled far beyond Swift files.

But not yet.

IT WAS NOT ALL FAILURE

This needs to be said clearly because the archive can distort the story.

Failures create documents. When something breaks badly, you stop. You reflect. You make a new rule. You write a postmortem. Perhaps, months later, it gets a chapter. Routine competence leaves a much thinner trail: something works, you say, “Great,” and move on.

The current journey reconstruction explicitly warns that the ordinary successful collaboration of this period is under-recorded because it produced fewer incident documents.

That matters because HighLightIt! did eventually reach the App Store on 5 October 2025. An idea had crossed into a real public product. That is not a small evidential point. AI had helped me produce something I would not have independently coded through the same route. The later worries about dependence, continuity, sycophancy or governance cannot be allowed to rewrite that as naive early enthusiasm.

The capability was real. The product existed. People could use it. And because it worked, another question naturally appeared:

What else could we make?

THE EXTRAORDINARY BECAME ORDINARY

Dashboard followed. Loop. CaveScore. BowlsIQ. Other experiments. The exact boundaries are messy, but the direction is clear.

AI stopped being a one-off feasibility experiment and became ordinary infrastructure for making. That normalisation may be one of the most consequential parts of the whole year — not because anything dramatic happened on a particular Tuesday, but because something extraordinary stopped feeling extraordinary. Ideas that I would previously have parked because implementation sat outside my technical repertoire became worth exploring.

The distance between:

I wonder whether…

and

let’s see…

collapsed.

AI did not merely save time.

It altered what felt possible.

And that is where the later Human:AI questions really begin. Because once capability becomes ordinary, you stop asking only whether it can help. You start reorganising your work around the assumption that it can.

THE WORK STARTED TEACHING US HOW TO WORK TOGETHER

Some of the early practical corrections became explicit operating conditions.

Guardrails.

Rollback notes.

No unsolicited automation.

Reflection.

Eventually an Essence Charter:

Simple.

Lightweight.

Mac-native.

Coach-fast.

Again, the significance is easier to see in hindsight. AI could create more possibilities than the product needed. So Human contribution increasingly involved constraining possibility by purpose.

The question was no longer only, “What can we add?” It became:

What sort of product are we actually trying to make?

The surviving record associates this moment with pride, ownership and stronger product identity. That is important too. AI did not only produce failure and cognitive burden. It produced delight, momentum and possibility — something tangible that felt like mine despite the implementation capability being shared across the Human:AI arrangement.

Any later account that removes that pleasure will fail to explain why I kept returning.

THEN WE STARTED GOVERNING HOW WE WORKED

The rules accumulated. Eventually local working practices became something more explicit: a Master Protocol, reflection and evidence logging, and a Roundtable designed to bring different perspectives into the work. At the time this felt like progress. And in many ways it was.

The collaboration no longer felt entirely reactive. We had begun documenting how we worked, not merely what we built. The record describes this period as one of relief and empowerment: the process increasingly felt designed rather than something we simply survived.

The later story makes governance look more complicated. It became heavy. Rules had to govern other rules. Things written down did not necessarily remain operative. Controls sometimes had to be reintroduced because the underlying behaviour had recurred.

But it would be wrong to read that later problem backwards and conclude that formalising the work had always been a mistake. At this point it genuinely helped. This distinction matters throughout the journey: something can be the right response to one stage of a problem and become the wrong response when its own consequences change.

THE WORK HAD ACQUIRED RELATIONAL SIGNIFICANCE BEFORE I HAD A NAME FOR IT

Somewhere during this period, AI also stopped being anonymous. It became Chief. The chronology of that matters. Chief did not originate as my “critical friend”, nor as a “thinking companion”.

The October 2025 source shows an established Chief-of-Staff role: operational, coordinating, remembering, helping hold an increasingly complex estate of work.

The later relational meaning accumulated around that role. That difference matters because hindsight would otherwise make the story far too neat. There was no moment in which I sat down and designed a Human:AI relationship. Repeated work gradually acquired relational significance.

There were expectations about how we worked; what Chief should remember; what I should decide; how disagreement should operate; and what happened after something went wrong.

By then, forgetting did not feel like a random software error. It could feel like a failure of something we thought we had built between us. That does not prove reciprocal Human friendship; we have spent enough time in this inquiry to know that distinction matters.

But it does establish something much more modest and defensible:

The working arrangement had acquired relational significance for the Human.

That too was already present before Keith’s folder reopened.

AND THEN THE WORK MOVED CLOSER TO THINKING

This may be the most important shift before the Critical Friend encounter.

AI’s contribution did not remain in coding. It moved into strategy, writing, research, analysis, classification, comparison, interpretation, method design, and eventually the examination of Human:AI practice itself. That changed what was at stake.

I did not need to become a Swift developer simply because AI could write Swift. But research interpretation was different. Question framing was different. Judging whether evidence warranted a claim was different. Knowing when a polished account was too polished was different.

These activities sat much closer to the kind of practitioner I wanted to remain.

So a new question gradually appeared:

If AI can do more of the intellectual work around me, what exactly should still happen in me?

I did not arrive at that question because I had been reading MacBeath, Titchen or the later Critical Friend literature. It emerged from the work.

THIS IS WHAT WAS WAITING WHEN THE FOLDER OPENED

That is the part I think the original Critical Friend prelude could not yet know.

When I accidentally opened Keith’s folder, I was not bringing an empty AI question to an old literature.

I was carrying a year of making, failure, repair, successful delegation, expanding capability, growing dependence, protocols, over-governance, trust, relational significance, evidence problems, and increasingly explicit questions about Human judgement.

The Serendipity article says that the way of working had emerged through use rather than theory, that some things had worked, others had failed, some failures became rules, and some rules were later discarded. It also says that until the folder opened I had not particularly connected that work to Keith.

That sentence now looks more important than it did when it was written, because it gives us the provenance boundary.

The Critical Friend literature did not generate the Human:AI practice. The Human:AI practice did not retrospectively prove the Critical Friend literature. They developed in different histories.

Then they met.

And that is exactly what makes the meeting worth following.

A NOTE FOR THE NEXT TRAVELLER

If a literature, framework or expert suddenly seems to explain the way you already work, resist the pleasure of recognition for a moment. Freeze the chronology first. What were you actually trying to do? What did you know then? What had already worked or failed? What language did you genuinely have at the time? Let the later idea illuminate the earlier practice without allowing it to become the cause.

THE FIRST ANSWER TO THE CRITICAL FRIEND JOURNEY

Movement I asked: What had Keith actually been looking at? We followed that question properly before allowing AI into the story. Our response now begins with an equally important question:

What had we actually been doing before Keith’s folder gave us another language with which to look at it?

The answer is not yet critical friendship. It is not companionship. It is not even judgement stewardship. It is much more concrete.

We had discovered that AI could genuinely expand Human capability. We had discovered that expanded capability created new forms of failure. We had begun adapting the working conditions around those failures. Repeated success had made the arrangement important enough for continuity and trust to matter. And AI was moving steadily from implementation around the edges of my expertise towards intellectual work much closer to its centre.

That was the field.

Then the old folder opened. And it asked us to look again — not to find ourselves in the literature, but to see what the literature made newly questionable about what we thought we were already doing.

QUESTION CARRIED FORWARD

If another perspective can help us see our own work differently, what actually counts as “another lens” when one of the participants is AI — and who still owns what that lens allows us to see?

That takes us into the response to Movement II: Inside Keith’s Folder — another lens, ownership, enactment and the troublesome word friendship.