A year of shipping software by writing specifications instead of code taught me a lot of things, but the one that surprised me most was not about the tools. It was about people. Spec-driven development, where you write a precise specification and an agent builds from it while a human still reviews every merge, is a real and teachable skill. It is not magic, and it is not a demo. A single person can move a product idea all the way into production and leave behind a system that makes the next effort faster. That is delivery.
The surprise is what the skill is actually made of, and how few people start with all of it.
So this is a map of that skill: the handful of things it rewards, who is already positioned to do it well, and how any role grows into the work. I am writing it in the open because the shortest version of the lesson is that the barrier is not talent. It is reps.
Four Skills, and Nobody Has All of Them
The work rewards a blend of four things: depth in a domain, technical and analytical chops, fluency with AI and the tooling around it, and the social skill to move people. Almost nobody arrives with all four.
Start with the one everyone is missing. AI as a way of working is new to all of us, so that is at least one skill every single person is short on to start, no exceptions. Layer on the usual gaps and most people land about three short: AI plus two more. A deep domain expert may have never touched the tooling. A strong engineer may lack the product depth to decide what is worth building in the first place. None of these are walls. They are learnable, and on the AI one, everyone starts from the same line.
That last part matters more than it sounds. In most skill shifts, the people who already know the new thing have a head start measured in years. Here the newest and most important input is one that nobody had a year ago. The field is more level than it looks.
Match the Person to the Surface
The most useful thing I learned is that different backgrounds are strongest on different parts of the work, and the move is to aim your existing strength where it pays off first.
Engineers are strongest on new features, where architecture and systems thinking are the hard part, so a greenfield build plays to their high. Data scientists shine on the back end of existing features, where the logic and the math live, and where knowing that a number measured two ways is not the same number is worth more than it looks. Designers own the front end of existing features, where the user-facing surface is everything and thinking in exact states and edge cases is the whole job. And product is the interesting open question, the connective tissue that stitches the other three together and a frontier nobody has fully mapped yet.
Play to your high first. It builds the confidence, and the evidence, you need before you go spend time on your low.
Which Way to Grow, and It Is Not the Opposite
The obvious move once you know your weak spot is to chase the exact opposite skill. That is usually wrong. Closing your low means picking the highest-leverage adjacent skill, not the most distant one.
For someone strong on the back end, a little front-end range helps, but architecture is often the bigger unlock. Broadly, engineers, data scientists, and designers should all reach toward the product skills of domain depth and owning the customer and the communication, because that is the part the agent cannot supply, and the part that decides whether the right thing gets built at all. Data science and design should also pick up more architecture, and a bit of each other's worlds.
Product managers grow in the other direction. They already have the domain and the communication. Their path leans hard into technical depth and architecture, plus picking one specialty, front end or back end, and going far enough into it to steer the machine with real judgment rather than vibes.
The Product Manager's Edge Is Social
That leaves an honest question about my own role, because on paper the agent is coming for the technical core of every job above. Engineering, design, and data science can all now lean on the agent for the build, and the gap between those disciplines narrows every month.
What the agent cannot do is the part that was never technical. It cannot build trust, read a room, negotiate a sign-off, align the people who own the code, or tell the story so the outcome actually lands with the people who decide what matters. That is the product manager's edge, and it is durable in a way the typing was not.
An edge is not a free pass, though, so here is where product honestly sits today. The role needs three things to run this way of working well: communication, domain knowledge, and technical knowledge. Communication is usually good. Domain is often only okay. Technical is usually low. And AI is new to everyone. It does not work if any one of those is weak, because every one of them has to be good or great. So the path is clear. Hold the communication edge, deepen the domain past okay, and close the technical and AI gap until all four clear the bar.
The Path Is Reps, Not Talent
None of this requires being the person who was always going to be good at it. The moves are nameable and repeatable. Hand the agent a real specification instead of a vibe. Make it list its assumptions and name what it does not know. Interview it before it writes a line. Keep the work isolated on a branch. Run an audit before you push. Save every correction as a rule so the same mistake cannot recur.
Start with one small, real change and do exactly that. Then do it again. Do it five times and the sixth is second nature. What accrues is not a one-time trick. Every correction becomes a memory, a gate, or a reusable skill, and the system you steer after a few projects is measurably better than the one you started with.
That is the real case for working this way, and it is why I am optimistic rather than nervous about it. The skill is learnable, the most important input is new to everyone at once, and the barrier that is left is reps and a willingness to be corrected in the open. I would rather be early and honest about that than sell anyone the magic version.