For twenty-five years I have been doing one thing under different titles. Web developer, agency owner, operations director, partner, founder, builder. The job descriptions changed. The actual work did not.
The actual work was always this: before anything could be built, somebody had to figure out what it was supposed to be. That somebody was usually me, or one of the people I had hired specifically because they could do it. We called it scoping, briefing, requirements gathering, specification, discovery, kickoff, planning. The names changed. The activity was the same.
A client would arrive with an idea. Sometimes a sharp one. More often a vague one wearing the costume of a sharp one. "We want a campaign site that converts". "We need a platform that does what our competitor does, but better". "We want an interactive piece that goes viral". "We need a customer portal". The brief, on the surface, sounded like instruction. Underneath, it was almost always a question the client had not finished asking themselves.
The job, before anything else, was to finish the question.
What is the actual outcome you are trying to produce. Who is the user, in real terms, not in marketing terms. What does success look like, measurably. What is in scope. What is explicitly not in scope. What are the constraints. What are the trade-offs you are willing to make. What is the smallest version of this that proves the concept. What would have to be true for this to be a waste of time. What would have to be true for this to be a hit.
Most clients did not arrive with answers to those questions. Many of them had never been asked them in that order. The conversation often took longer than the client expected, because they had assumed the project was about execution. The project was almost never about execution. The project was about specification, and execution was the easy part once the specification was real.
This was true at every scale. It was true for the small SME website where the client was the founder and we had to ask them what their actual customer looked like. It was true for the Volkswagen one-million-users campaign where we had to specify the technical limits of the system before we could commit to a launch. It was true for the McDonald's hardware piece where the constraint was a physical placemat. It was true for the European Schweppes platform where the constraint was nine markets with nine sets of legal requirements. It was true for the small ones and it was true for the prestigious ones. The size of the project changed the stakes. It did not change the discipline.
What I now understand, with hindsight, is that this skill had a name nobody used. We called it experience. We called it judgment. We called it being good at the work. None of those words quite captured what it actually was, which was the trained ability to take a fuzzy goal and produce a clear specification of the work required to reach it. We were doing the activity all day, every day, for twenty-five years, and we never had a clean word for it.
The word turned out to matter.
A native of a country I had not realized I lived in
When generative AI tools became genuinely capable, I started building again. The first time I sat down with a model and tried to produce something real, I expected the experience to be foreign. The languages I knew were old. The frameworks had moved on. The defaults I remembered were embarrassing. I expected to be a tourist.
I was not a tourist. I was a native of a country I had not realized I lived in.
The reason was that the model did not need me to know modern syntax. It produced modern syntax on demand. It did not need me to remember how to set up a build pipeline. It set up the build pipeline if I asked it to. The thing the model needed from me, the thing it could not do without, was the specification of what was being built. That was the part where my agency years suddenly mattered, and mattered far more than I had expected.
Every conversation with the model was, structurally, the same conversation I had been having with clients for twenty-five years. What is the actual outcome. Who is the user. What does success look like. What is in scope. What is not in scope. What are the constraints. What are the trade-offs. The model, like a junior developer, would happily produce something for any prompt. The quality of the output depended almost entirely on the quality of the prompt, which depended on the quality of the thinking that produced the prompt, which depended on whether the person typing the prompt knew how to specify a thing.
I knew how to specify a thing. I had been doing it for a quarter of a century. Most of the discomfort I had expected - the feeling of being behind, of being too old for the new tools - never arrived. The tools had moved my old job into a different shape, but the job underneath the tools was the same job.
This was not a comforting realization. It was a strange one. I had assumed, like many founders my age, that the new technology would require new skills, and that the skills I had spent decades developing might be partially obsolete. The opposite turned out to be true. The skills I had developed were not partially obsolete. They were partially hidden. The tools had been hiding them by absorbing the easy parts of the work - the syntax, the boilerplate, the implementation details - and exposing the harder parts that had always been the actual job.
The actual job, with the implementation details removed, was specification.
The instinct AI rewards
The argument I want to make is broader than my own experience. It is that the people who will use AI well, in any field, are disproportionately the people who already had this instinct before AI existed.
Not because they are smarter. Not because they prompt better. Because the underlying discipline - the ability to take a fuzzy goal and turn it into a clear, specific, scoped description of what needs to happen - is the same discipline AI rewards. It was always the bottleneck of good work. AI did not introduce that bottleneck. AI made it visible by removing everything around it.
You can watch this happen now in any field where AI tools have become competent. The people producing great output are not the people with the best prompts. They are the people who know what they want before they sit down. Their prompts are usually short, because their internal specification is dense. They have done the thinking. The model just produces the result.
The people producing weak output are usually skipping the same step. They sit down with the model, type something general, and then iterate against whatever the model produces, trying to nudge it toward something better without ever fully specifying what better would look like. The conversation gets longer. The output gets prettier without getting clearer. Eventually they settle for what they got, because settling is easier than going back to the specification work they should have done in the first place.
This is the same pattern I watched in clients for twenty-five years. The clients who got the best work from us were the ones who could brief well, or who were willing to let us run a real specification process before any work began. The clients who got worse work were the ones who treated specification as overhead and execution as the real work. They lost time in revisions. They lost money in scope creep. They lost quality in compromises. The cost of skipping the upfront thinking always showed up later, in worse outcomes, every time. AI has compressed the timeline of that lesson. The cost of skipping shows up not in three months, but in the next prompt cycle.
Specification fluency
There is a practical implication of this argument that has changed how I think about my work, and increasingly how Peter and I talk about agency leadership.
The skill that makes someone valuable in an AI-augmented world is not technical fluency. It is specification fluency. It is the ability to take a problem, ask the right questions, produce a clear scope, define what good looks like, identify the constraints, articulate the trade-offs, and write all of that down in a form a capable producer - human or model - can act on without ambiguity.
This skill is acquired the same way every other operator skill is acquired. By doing it under pressure, with feedback, for years, on real problems that have real consequences. There is no shortcut. The instinct that lets you read a brief and immediately see what is missing, what is contradictory, what is unspecified, what will cause problems three weeks in - that instinct comes from having been wrong about all of those things, in public, in front of clients, for a long time.
Agency operators have been training this instinct for their entire careers. So have product managers. So have engineering leads. So have consultants, lawyers, doctors, anyone whose job involved turning client uncertainty into actionable clarity. The skill was always there. The tools were just hiding how much of the work was actually about it.
The interesting question for anyone in those roles now is whether they recognise what they have. Many do not. Many are still treating AI as a tool problem, looking for the right course, the right prompt library, the right framework. The tool is not the problem. The instinct they already have is the answer. They have not noticed it because nothing was forcing them to look.
For founders running businesses where this instinct lives in their head, there is a second implication. The instinct is one of the things you carry. It is also one of the harder things to scale, precisely because it is mostly tacit, mostly developed through experience, and mostly invisible until it goes missing. Building an agency that runs without you means, among other things, building an agency where this instinct lives in your team, your documents, your processes - not only in your nervous system.
What the four years were waiting for
I am writing this article partly for myself. I spent four years out of digital work, partly because I thought my skills were aging out of relevance. They were not. They were waiting for the right tools to make them legible again. If I had known that at the time, I would have spent less of those four years assuming I was behind, and more of them building.
I am also writing this for the operators who are still partly hidden from their own value. If you have spent a decade or more turning fuzzy briefs into clear work, if you have an instinct for what to ask before anything gets made, if you can sit in a kickoff and feel the missing pieces of a brief before they cause problems - you already have the central skill of this era. The tools have done you a favor you may not have noticed. They have removed the parts of the work that used to obscure what you were actually good at.
The competition is not what most people think it is. The competition is not people who prompt better. The competition is people who know what they want and can describe it precisely.
That, it turns out, is what twenty-five years of agency work was teaching us all along.
We just did not have a name for it.
Originally published on LinkedIn.