On this page
Questions you could ask me
Questions you could ask me
Could the same questions matter in your work?
Or describe your own challengeI’ve spent a substantial part of the past few years working with generative AI. On many days that has meant ten hours or more: experimenting, building, testing, investigating failures and returning to problems that weren’t yet resolved. By rough arithmetic, the total is now in the region of 10,000 hours.
I haven’t kept a stopwatch running, and I don’t regard 10,000 hours as a certificate of expertise. What interests me is what that amount of sustained work has changed: not simply what I can make with AI, but how I judge what it produces.
I started with something familiar
One early application was Use Up Food, an ingredient-led recipe generator that produced a structured recipe before generating its accompanying food photograph. It wasn’t a grand commercial plan. I wanted to understand the technology by building something people could immediately recognise and evaluate.
The challenge wasn’t just getting AI to return a recipe or an attractive image. It was making different recipes readable in a consistent format and giving their photographs something like a shared studio appearance. The results weren’t perfectly uniform, but the effort revealed the difference between an impressive individual response and a repeatable product experience.
That application also gave me a way to demonstrate what I was learning. It contributed to conversations that led into my AI-related product work at Zegeba, where the questions became more directly connected to professional design and delivery.
From asking questions to directing work
My earlier career had already involved more than thirty years of software delivery: understanding needs, designing systems, writing and reviewing implementations, dealing with constraints and seeing the consequences of decisions over time.
AI didn’t supply that experience. It changed what I could do with it.
I began to move from asking for a particular answer to directing connected pieces of work: interface design, application behaviour, architectural decisions, testing, documentation and the relationships between them. A working result could be examined sooner, and I could move between disciplines while keeping the original intention in view.
That doesn’t mean one person can replace every specialist. Domain expertise, independent review, collaboration and operational accountability continue to matter. But the distance between an idea and something that can be tested has changed considerably.
The value of knowing what not to do
At first, progress was easy to confuse with the quantity of output. A useful-looking response, a fresh implementation or a long explanation can create the impression that a problem is moving towards resolution.
With experience, I’ve become much more interested in whether the work is actually converging on the original requirement. A proposed architectural improvement may introduce unnecessary complexity. A plausible explanation may be based on a file that nobody has inspected. A correction may address a symptom while leaving its cause untouched.
Knowing when not to accept an answer, when not to add a feature and when not to change an architecture has become as important as knowing what to ask for.
Learning to recognise the rabbit holes
Some of my most valuable lessons have also been among the most frustrating.
I’ve spent days pursuing problems that appeared to be close to resolution. A proposed correction would look convincing, only to reveal another complication. Another correction would follow, then another, until I found myself considerably further from the original objective.
I’ve learned to recognise those situations much earlier.
There are patterns: repeated fixes that don’t address the underlying cause, assumptions presented as facts, unnecessary architectural changes, and solutions that become progressively more complicated without producing a better result.
Earlier in my experience, I might have continued exploring those possibilities. Today, I’m much more likely to recognise the direction of travel, stop the process and return to what can actually be established.
Sometimes that means examining the original requirement. Sometimes it means tracing the implementation, testing an assumption or rejecting an approach entirely.
The hours spent discovering these lessons were often frustrating. But they weren’t wasted. They’ve changed the way I work.
Experience doesn’t necessarily prevent you from encountering a rabbit hole. It helps you recognise one before you’ve spent days exploring it.
Evidence before confidence
My working discipline has become much more explicit: OBSERVE → TRACE → PROVE → CHANGE → TEST. Begin with what the system actually does. Trace the behaviour to its implementation. Establish a cause before proposing a correction. Make the smallest justified change, and test whether it improved the result.
That sequence isn’t a claim that every investigation is simple or linear. It’s a guard against treating a convincing AI explanation as if it were evidence. When an assistant is asked to work on an existing codebase, the codebase must be the authority. If it has not been inspected, its architecture cannot be described as known.
I now try to make decisions traceable, reduce the number of assumptions left unstated and keep control of what an AI assistant is authorised to decide. Sometimes the right result is a refusal to proceed without evidence.
The product reflects the method
That discipline has influenced my product work, particularly determinIQ, the governed data intelligence platform I’ve been developing. Its analytical results come from deterministic processes; AI is used for wording and interpretation of intent, not as an authority for the calculations.
There is a connection between how I try to direct AI during engineering and how I’ve designed the product to constrain it. In both cases, an answer is only as useful as the evidence and authority behind it. A confident statement isn’t a substitute for a verified result.
That doesn’t prove the product is complete or remove the need for testing. It describes the design principles I have chosen and why they matter to me.
What the hours have given me
The greatest improvement in my use of AI hasn’t been learning how to get answers faster. It’s learning how to recognise the answers that aren’t taking me anywhere.
The early experiments taught me what the technology could produce. Professional work exposed more demanding constraints. Building larger systems taught me to coordinate work without surrendering architectural responsibility. The failures taught me to spot unproductive patterns sooner.
I remain curious about what is possible, and I still encounter unexpected problems. The difference is that I no longer treat every possible direction as worth following.
Ten thousand hours is a convenient way to describe the scale of the experience, not its value. The more meaningful measure is the judgement accumulated along the way — and what that judgement allows me to build, question and improve.
Research context
Research programme: From intention to implementation: human-directed AI-assisted engineering
Topic: AI-assisted engineering
Explore further
Further reading is listed below.