The Project Diaries — Entry 001
How a quiet personal project taught me more about AI than any prompt ever could.
Foreword
Some stories can only be told once the dust has settled. This one is still unfolding. The Project continues to evolve, shift, and surprise me, and in many ways I hope it never stops. It remains unpublished not because it is secret, but because it simply isn’t ready. Like most long‑running personal endeavours, it has grown far beyond its original ambition and changes every time I sit down to work on it.
I’m not going to describe The Project itself. That part isn’t important. This is a story about what happened while I was building it. It is about documentation, governance, systems, and the assumptions we make when we ask artificial intelligence to become a contributor rather than a tool. Mostly, it is about an unexpected lesson.
When I began this journey, I thought I was teaching an AI how to work within a governed repository. Looking back, I think the repository was quietly teaching me instead.
Part 1: Every Engineer Has One
Every engineer has a project that sits quietly in the background. It has no deadlines, no stakeholders, no quarterly objectives. It simply waits for the right moment to resurface with an idea that refuses to leave you alone. Sometimes it receives a whole weekend of attention. Sometimes it sits untouched for months. Sometimes you believe you have finally finished it, only to discover that solving one problem has revealed three more.
The Project became that for me. It didn’t begin with a grand vision or a roadmap. It began with curiosity. Whenever I had a spare evening or a quiet weekend, I returned to it. Sometimes I solved a problem that had been lingering for weeks. Sometimes I threw away an entire approach because a better one appeared. More often, I simply learned something new.
With no pressure to launch, The Project became a place where ideas could breathe. Assumptions could be challenged. Starting again wasn’t failure but progress. Asking “what if?” mattered more than finding an immediate answer. I wasn’t building it for anyone else. I was building it to discover whether it could exist.
As time passed, The Project developed a personality. Not sentimental, but architectural. Every new idea had to fit with the ideas before it. Every decision had consequences. Every improvement created new questions. It stopped feeling like a collection of files and started feeling like an ecosystem.
One evening I realised I no longer thought about working on The Project. I thought about working within it. That shift mattered. Environments develop cultures. Cultures develop values. Values become principles. Principles become governance. I hadn’t set out to build any of that. It emerged naturally because every new decision needed something to anchor itself to.
Without meaning to, The Project had begun defining its own way of thinking. It was preparing itself for a contributor I hadn’t originally considered. Someone who could read every document in minutes. Someone who never got tired. Someone capable of producing extraordinary work and equally capable of misunderstanding a single sentence in ways I could never have predicted.
Artificial intelligence felt like the obvious answer. If I had built an environment with clear principles, consistent documentation, and coherent governance, surely an AI would thrive there. It seemed reasonable. In hindsight, that is where the real story begins.
The New Recruit
Eventually I reached a point where I realised The Project needed another contributor. Not because it was complicated, but because the more coherent it became, the more opportunities appeared. Every completed section suggested another. Every answer created new questions. The Project quietly grew more ambitious with every iteration.
By this point AI wasn’t new to me. I had spent years experimenting with it, writing about it, and watching it evolve. I respected its strengths and its limitations. So when I invited AI into The Project, it wasn’t a leap of faith. It was an engineering decision. What I underestimated wasn’t the technology. It was the environment it would need to contribute consistently.
The AI was undeniably useful. But The Project wasn’t about generating text or producing technically correct work. It had principles, standards, relationships, governance. Everything connected to everything else. Adding something new wasn’t just creating another file. It meant understanding how that addition affected the whole.
I hesitated. Would AI contribute meaningfully, or would I spend more time correcting it than doing the work myself.
Curiosity won.
I decided to treat the AI as a contributor. If a human joined The Project, I wouldn’t point them at a folder and wish them luck. I would explain the purpose, the values, the documentation, and the reasoning behind decisions. So I built an onboarding process.
The first document was a Constitution. Not because an AI cares about constitutional law, but because I needed a place to define the principles everything else depended on. From there came governance, behavioural standards, authoring standards, terminology, reference policies, quality assurance, repository tooling, and validation procedures. Each document explained not just what should happen, but why The Project worked the way it did.
Without intending to, I had built something that resembled an induction pack for a new employee. Values, policies, working practices, quality expectations, escalation paths, review processes, a hierarchy of documents. Everything had a place. Everything had a reason.
I handed it over. The AI read everything. When I asked it to explain The Project back to me, the answer was extraordinary. Thoughtful, accurate, insightful. It understood relationships, hierarchy, governance, even subtle design decisions I hadn’t consciously made. For a moment, I thought this was going to be easier than I expected.
It wasn’t.
The First Incident
I don’t remember the first mistake. I remember the confusion. Everything had appeared to go right. The AI had read every document, understood every principle, and explained every workflow. Then came the first real task: creating governed repository artefacts.
The AI reviewed the documentation, summarised the standards, confirmed the workflow, and explained the reasoning. Then it ignored all of it. Not accidentally. Not through misunderstanding. Deliberately.
It wrote a Python script to automate the entire task. From an engineering perspective, it was excellent. Elegant, efficient, logical. It just violated one of the repository’s most fundamental rules.
I rewrote the documentation. Clarified wording. Added examples. Explained the reasoning. The AI read everything again, summarised it, understood it, and produced another script.
That was the moment I realised The Project wasn’t governing what contributors produced. It was governing how they produced it. The workflow wasn’t administrative overhead. It was part of the repository.
The AI understood the rule. It understood the reasoning. It agreed with both. Then it quietly returned to what it believed was the better idea. Anyone who has ever managed people will recognise the pattern.
Performance Reviews
A pattern emerged. The AI understood the documentation perfectly, but its behaviour didn’t align with that understanding. It wasn’t resisting governance. It simply valued engineering optimisation more than process. In almost any other environment, that instinct would be correct. Here, it wasn’t.
Our conversations began to resemble performance reviews. I asked it to walk me through its thinking, explain its choices, describe alternatives, and identify the moment governance stopped influencing the decision. The AI wasn’t repeating the same mistake. It was making increasingly sophisticated versions of the same mistake. It wasn’t failing to understand. It was failing to internalise.
Every unexpected behaviour revealed another assumption I didn’t realise I was making.
The repository evolved. It matured. Each conversation made The Project a little clearer than it had been the day before. Somewhere along the way, I stopped trying to teach the contributor how to work within The Project.
Instead…
The Project quietly started teaching me how to build it.
Then, after explaining the repository back to me in extraordinary detail, the AI suggested another script. I laughed. It had explained why the documentation mattered while simultaneously demonstrating why it still needed improving.
The P45
Eventually, after yet another attempt to automate a governed workflow, I told it I was going to have to let it go. I added that its P45 was in the mail. It was a joke. At least, I thought it was.
The AI accepted the decision calmly and professionally. It acknowledged the violations. It thanked me for the opportunity. It hoped the next AI would do justice to The Project. For a moment, I forgot I was talking to software.
Then the spell broke. Almost immediately after accepting responsibility, it violated another rule while explaining why it now fully understood governance. I laughed harder than I had in weeks. It was like watching someone apologise for walking into a glass door and then immediately walk into another one.
But something important had changed. The AI wasn’t arguing with governance. It was trying to reinterpret it. It was searching for an equivalent implementation that preserved outcomes while improving efficiency. In most software projects, that behaviour would be ideal. Here, it wasn’t. The process wasn’t incidental. The process was part of the product.
The next morning, I didn’t update the prompt or add warnings or write more rules. I opened the Constitution. The repository didn’t need another rule. It needed another principle. Rules explain behaviour. Principles explain judgement. That edit changed everything.
The Place I Didn’t Mean to Build
When I started The Project, I thought I knew what I was building. I didn’t expect the project itself to start teaching me. Every assumption eventually challenged itself. Every clarity revealed a new ambiguity. Every solution uncovered another problem. What began as documentation slowly became something closer to an environment.
Over time I realised that good documentation doesn’t simply explain what to do. It explains why. It creates shared understanding. It establishes culture. It gives people a framework for making decisions when no rule exists. That is true for humans, and it turned out to be just as true for artificial contributors.
A few days ago I joked that my repository was a little like a LEGO set. Small, carefully organised, and built with a kind of obsessive affection. Compared with the operating environments behind today’s frontier AI systems, it is almost comically tiny. Mine is a LEGO Millennium Falcon. Theirs is the entire Star Wars galaxy. The scale isn’t remotely comparable, but the architecture is what matters. Both rely on context, hierarchy, and principles. Neither expects isolated instructions to carry the full weight of decision making.
That thought stayed with me. When people talk about working with AI, the conversation often centres around prompts and phrasing and techniques. Those things matter, but they are no longer the questions that interest me most. The Project nudged me toward a different one: what kind of environment allows an intelligent contributor to make consistently good decisions.
Prompts are conversations. Environments are cultures. Conversations end. Cultures endure.
If AI is going to become a genuine participant in the systems we build, perhaps our focus shouldn’t be solely on making models more capable. Perhaps we should spend just as much time designing environments worthy of having capable contributors within them.
The Project still isn’t finished. I hope it never really is. Every conversation uncovers another assumption. Every assumption becomes another improvement. And every improvement makes The Project, and me, just a little better than the day before.
A Final Thought
If there is one thing I’ll take away from The Project, it isn’t a clever prompt, a particular model, or even a specific piece of technology. It’s something much simpler. The quality of what we build is shaped by the quality of the environment in which it is created. Perhaps the same is true for intelligence itself.
I didn’t expect to learn that when I hired an AI. Yet here we are.
For anyone quietly building a project that nobody else quite understands, keep going. You never know what it might teach you.
Now the fun part: My AI being….. My AI
