All notes

AI took the training ground. Craft is learned on purpose now

Last week I wrote that building alone got loud — no teammates, no hallway critique, just you and a machine that agrees with you. There is a second missing piece under that one, and Designer Fund put numbers on it: in their AI in Design 2026 survey of 900+ designers, half have pushed AI-generated code to production, and one of the loudest worries in the write-in responses is craft atrophy and skill development.

Here is my read on why that worry is structural, not nostalgic.

AI didn't just speed up the work. It automated the exact work that used to train people.

On this page

What got lost when the bottom rung disappeared?

Think back to how the last generation of designers and engineers actually learned. Not from courses — from the tedious stripe at the bottom of the project: hand-coding the 14th card variant, fixing contrast on a dark-mode palette, chasing a three-pixel spacing mismatch, reading a 900-line file to change one button. Nobody liked this work. It is where pattern recognition got built anyway.

That stripe is now a prompt. The agent does it in seconds and nobody learns anything from it. The craft ladder did not get shorter at the top — it lost its bottom rungs, and those are the ones you climb first.

Why can't you audit output you never learned to make?

This is where the risk stops being theoretical. Craft fundamentals are not for the machine — they are the reason your own judgment works. If you have never hand-fixed contrast, you cannot feel when a generated palette is eight percent off on dark backgrounds. You cannot distinguish a good generated layout from a merely plausible one, so plausible wins, and plausible is where the average lives.

METR's randomized trial made the perception problem measurable: senior open-source developers estimated AI made them ~20% faster when it actually made them 19% slower. Your felt speed is unreliable. For experienced people that is annoying. For people who never built fundamentals, it is worse — they have no independent way to check the agent's work, so the audit step silently vanishes and plausible code ships straight to production. Half the surveyed designers already do this. The other half are doing the checking, and craft is what that checking is made of.

How do you learn craft on purpose in 2026?

Nobody immerses you in the grind anymore by default, so schedule it against the current:

Run build sessions and ship sessions. In ship sessions, use the agents at full speed. In build sessions, close them. One small, boring change, by hand.

Hand-make the worst jobs on purpose. Fix contrast yourself. Restyle one component against your tokens with no generated scaffolding. It is the same content the apprenticeship used to hide inside, just now you have to go get it.

Read unfamiliar code end to end. Give the agent the task, then read what it wrote like a reviewer with authority, not a tourist. If you cannot say why a line is wrong, that is your next training rep.

Do one thing slower than the model. Deliberate slowness is the only pathway left to pattern recognition that hits fast without supervision.

I do this on my own builds and it is annoying every time, which is how I know it is working.

Is craft atrophy real, or gatekeeping paranoia?

Real, and the survey is the least of it — the mechanism is arithmetic. Skill acquisition needs repetition on doable-but-hard problems. AI removes exactly those problems. What you get is a workforce that can ship and cannot judge, which is precisely the state where AI products end up generic: someone has to feel the miss before anyone can fix it.

The fix is not less AI. It is choosing to learn on purpose in an environment where learning used to be a side effect and is now a decision.

Frequently asked questions

  • It is the missing stretch of work that used to teach juniors the fundamentals: hand-styling components, fixing contrast, matching spacing scales, reading unfamiliar code to make one small change. AI now does most of that work instantly, so many practitioners reach shipping velocity without ever building the judgment that made skilled people skilled.

About the author

mosh

mosh is a product designer for growth, working with design thinking and ever-improving design systems. What matters: fixing conversion, whether in B2B dashboards or direct-consumer apps.

Keep reading