I still write code by hand. You should keep a slice of your workflow that way too

AI coding tools are an easy yes. GitHub's Copilot trial had developers finish a greenfield task 55% faster, which is why it sold so fast. I use them on basically everything. And I still carve out a slice of every week where I write code with no AI at all — small scripts, my own tools, things no one will see. That slice is not nostalgia. It is maintenance on the thing I complain about the most: taste.

The reason lives in a contradiction most people never hold in their head at once. METR's randomized controlled trial measured experienced open-source developers on mature codebases and found they got 19% slower with AI allowed, after predicting a 24% speedup. They were wrong before and after — but not lying. Faster generation does not mean faster decisions, the same trap I described in The last 20% is the whole point. The last stretch of work is decision, not typing, and no model speeds that up.

Why do I still write code by hand?

Two controlled experiments, opposite results, and both are true. GitHub timed a clean new task: 55% faster. METR measured an experienced person on a codebase they know deeply: 19% slower. The tool did not change. The context did. AI is phenomenal at greenfield reconstruction and genuinely marginal at mature judgment — and, most tellingly, it did not help the people who already had the context. That sentence is the entire story.

The people who slow down are the ones who stopped being the author. When the model does the work, you stop making the small, boring, exact decisions that built your skill in the first place.

What do the two speed studies actually say?

You have to sit in both numbers so you do not grab the one that agrees with you. GitHub timed 95 developers building an HTTP server in JavaScript. The treatment group finished 55% faster with Copilot. METR ran 16 contributors through 246 issues in large repositories they were responsible for. With AI allowed, they took 19% more time. On something you have done a hundred times, the model's suggestions cost you motion instead of saving it, because you already had the answer in your hands.

That is the piece almost everyone skips. The slowdown is not about weak models — the model is weakest at exactly the part you are already good at. Use AI for the part you are bad at, and reserve the part you know cold.

How does AI coding erode taste?

Taste has a source, and the source is a calibration loop: you decide, you watch it fail, you fix it, and next time your instinct is a second better. A decade of that builds a library of "that's off" gut calls. AI cuts the loop in two places at once. It produces the code, and it also stops you from ever shipping the wrong first version that was the reward.

That is the same mechanism behind Your AI eval set is your taste, made measurable. Once the machine generates everything, you never put a wrong thing in front of a real user yourself, so your internal standard stops getting calibrated. Hand-building is that feedback loop, deliberately protected.

What does a hand-build window look like?

It is small and unambitious, and that is the point. I set aside about 20% of my coding time where I write zero generated code. Real examples from my week: a tiny CLI I will probably throw away, a layout bug fixed with long handwritten CSS, a script that renames files.

I do it even though the AI would finish in two minutes, because the intent is not the output. Debug one thing yourself. Let the compiler bite you once. Design an import and feel it smell before a test does. Those are the repetitions that keep your calibration sharp, and you cannot get them by reviewing.

Does hand-writing actually cost you time?

Less than you think, and it pays a dividend. When AI hits the part of your codebase it cannot understand — the area with 1,500 commits of context — the person who still reads code by hand knows it instantly. They spot the wrong default, the N+1 query, the state one race away from a bug. The person who only reviews and accepts AI output has no instincts to bring.

The engineers who stay fast over years are not the ones who let a model write every line. They are the ones who kept writing, so they kept their judgment. The skill you refuse to delete is the one AI is worst at and the one that makes everything else faster. That is the hand-build slice — and it is cheaper than relearning the skill in five years.