If you have been writing software long enough, you remember a previous wave that was supposed to replace programmers. Each time, the craft changed — but it did not disappear. The question is not whether tools get better; it is which human skills still compound when the baseline rises.
Here are the ones I bet on for the rest of the 2020s.
Taste in product and UX
Anyone can generate a landing page. Few people can decide what to leave out, which customer pain is worth solving first, and how friction maps to trust. That is not a prompt engineering trick. It comes from watching real users struggle without defending your first idea.
Practice: run five moderated sessions a year, even on internal tools. Watch the recording without skipping to the happy path.
Systems thinking
Code lives inside networks, databases, caches, queues, and human processes. A model may propose a function; it takes a human to see how that function behaves when a replica is stale, when a queue backs up, or when an operator runs a migration at midnight.
Practice: after every incident, write a one-page timeline: what failed, what masked the failure, what you will instrument next. Patterns emerge faster than raw intuition.
Clear technical writing
Specs, RFCs, postmortems, and good commit messages scale your influence. AI can draft, but it cannot choose the audience, the scope, or the honest tradeoffs you will defend in a room.
Practice: rewrite one doc a week to be shorter. If stakeholders can skim the first screen and know the decision, you win.
Security and abuse instincts
Attack surfaces grow faster than feature lists. Tools help patch known issues; humans still have to imagine what a motivated stranger tries first.
Practice: threat-model one feature per quarter on paper — data flows, trust boundaries, what you log — before implementation freezes.
Kind, direct collaboration
Shipping is a team sport. Debugging ego, aligning priorities, and giving feedback that changes behavior without burning trust — models do not do that for you.
Practice: in reviews, separate "must fix" from "teaching moment," and never more than three must-fix items per round.
The through-line
Automation raises the floor. Judgement, narrative, and accountability raise the ceiling. Invest there and the tools become leverage instead of noise.