The Director's Seat — How I Work With AI
Every page on this site was touched by AI. I would rather explain the division of labour than have you guess at it.
The claim
I am an AI-assisted worker. Not in the euphemistic sense — "I use Copilot sometimes" — but structurally: AI executes most of my computation, drafts most of my prose, and tutors most of my learning. What I do is direct it. I choose the questions, set the standards, decide what gets claimed and what does not, audit the output, and carry the accountability when it is wrong.
I think of this as the director's seat. A film director does not operate the camera, act the roles, or score the soundtrack — and nobody says the film is not theirs. What makes it theirs is judgement applied at every decision point: what to shoot, what to cut, what the thing is for. That is the honest description of how this site was built, and increasingly, of how I work.
The plain version: I had problems worth solving and capability I did not have. Directing AI let me do more than I could have done alone. It did not let me do it without understanding it — and this page is about the difference.
What that looked like in practice
The bearing study started because I had a real decision in front of me — an escalator monitoring pilot at work — and no analysis capability of my own. So I directed an AI through it. I picked the dataset and the question (when a bearing dies, which measurement tells you first?), demanded the study separate what is validated from what is one bearing's worth of evidence from what is not claimed at all, and refused the additions that would have made it look more impressive — no retrofitted prediction model, no machine learning for sophistication's sake.
The AI wrote the code and ran the numbers. The epistemics — deciding what not to claim — were the part I would not delegate.
The learning came after, and separately. Directing an analysis you cannot audit is a trap; fluent output with uncertain grounding is how errors travel in confident packaging. So I ran a structured self-training sprint with an AI tutor: the vibration measurement chain upward, electrical signature analysis, industrial protocols, OT security. Not by reading — by deriving. I would reason my way to a mechanism and the AI's job was to tell me what the field already calls it.
Several things I "invented" turned out to be real: the wireless deployment model, acoustic beamforming, motor slip. I derived them wrong-named and got corrected. That loop — derive, get labelled, get corrected, drill until it survives pressure — is the fastest learning I have done, and it exists specifically so I can tell when the AI is wrong.
The failure record is part of the method. The sprint included cold-tested drills, a mock sales panel that went badly on purpose, and a cross-examination of this very website in which I disclosed exactly this division of labour rather than pretending to analyst-tier depth I do not have. My drill files document every fabrication I produced under pressure, mapped to its trigger. I keep them because auditing the machine starts with auditing the operator.
Why I work this way on purpose
There is a failure mode I take seriously enough to apply to myself: as AI briefing layers replace raw data, the people who could check the machine stop being able to read the numbers. The tool that makes expertise unnecessary quietly erases the expertise needed to catch it failing.
So the rule I run is: AI does the labour; I maintain the fluency to audit the labour.
- Prose drafted by AI gets the read-aloud test. If I would not say it with my own mouth, it gets resaid.
- Numbers get a source or they get cut.
- Claims get one filter: could I defend this line, in person, to a sceptic? If not, it does not ship.
There is a corollary I hold to, and it is the one that constrains this site most: the site must never outrun the author. When the cross-examination found a gap between what a page implied and what I could defend, the fix was to move the label — not to cram until the claim became true.
What I am claiming, and what I am not
Claiming. I can take a domain adjacent to mine, direct AI through real work in it, verify the output against understanding I built independently, and be precise in public about which layer is whose. The receipts are this site: the studies, the tools, the failure write-ups, and this page.
Not claiming. That I performed the computation. That AI-assisted means effortless. That direction substitutes for expertise — it does not, and it requires a specific expertise of its own: knowing what to ask, what standard to hold, and when the confident answer is wrong.
That last one is the skill I am building deliberately, and the one I think matters most in the next decade of technical work.
Where the line sits on each page
Every project carries a "How this was built" note stating what was mine and what was the machine's, because the split genuinely differs between them. The electrical work was specified, debugged and corrected by me against an AI that repeatedly produced confident nonsense. The bearing study was executed by AI under my direction on a question I chose. The retrieval-gate system was designed in plain language by me and implemented by AI, and the design decisions in it are the part worth reading.
Two of those are more mine than the third. Saying so costs nothing and is the whole point.
Analysis is reproducible from the linked repositories. Any line on this site, I will defend in person — and where I cannot, it says so.