Handwritten Code Isn't an Edge Anymore

For a couple of years I treated writing code by hand like a moral position.

In mid-2024 I wrote about catching myself alt-tabbing to Claude after about ten seconds of being stuck on a bug, and how it made me feel a bit ill. My answer was to sit with problems longer and do chunks of work with the AI tools closed, so I’d find out which muscles had gone weak before an outage did. As late as September 2025 I was still writing posts about how coding skills still matter.

I was one of those devs. I used the tools every day, but I believed the part where I typed the code myself was where my value lived. The AI was autocomplete with a personality, and the craft was mine.

That belief didn’t make it through 2025.

Anthropic put Claude Code out as a research preview in February alongside Claude 3.7 Sonnet. It went generally available with Claude 4 in May. By late November, Opus 4.5 was the first model to get past 80% on SWE-bench Verified. Somewhere in that stretch, and I can’t point to a single day, the agent stopped being something I checked on every thirty seconds and became something I handed work to.

It felt sudden. It wasn’t, really. The models had been getting better for years, but there’s a point where a tool goes from needing a babysitter to getting on with the job, and you notice it all at once.

Writing the code became the easier part. That was a hard thing for me to accept, because I’d spent well over a decade getting good at it.

The hard parts are the same as they always were. Working out what to build. Knowing a design is wrong before you’ve sunk a week into it. Reading a diff and spotting the line that’s going to cost a client money.

In August I spent a Sunday auditing the billing side of a client’s app and found a payment field that turned $125.40 into $12,540. The bug was one regex character class. Typing that regex was never the difficult bit. Knowing a money field should never multiply someone’s payment by a hundred, and going looking for the places where it might, took experience.

That’s why I don’t buy the idea that AI replaces developers. It augments them. Hand Opus 5.5 to someone who’s never shipped software and you’ll get a lot of code very quickly (some of it will even run). Hand it to someone with years of production scar tissue and they become superhuman. All that experience doesn’t go to waste. It’s what’s doing the steering.

I can point at my own year as proof.

In January I launched AskBad, a Q&A site where the worst answers win. In March I built a CLI to control my iZone air conditioning. May was SSR for Aurelia 2 and Tidewater, a Three.js water kit I pulled out of a game I’m making. July was an Instagram downloader for my Mac.

Then August got silly. Preset Bridge, RangeSearch and RangeMenu all went live within two weeks of each other, three products, each with its own site. And this morning I posted about the NAM-to-Proxy converter, which meant reverse engineering how the Helix Stadium runs its clones closely enough that my output matched Line 6’s down to rounding errors.

That’s on top of a day job, contract work, two kids and a guitar habit.

On 1 January I wrote a post about shipping more. I talked about my notes folder full of app ideas, the half-finished prototypes and the domains I’d bought at 2am in a fit of optimism. I meant every word, but I’d made that promise to myself plenty of times before. This time I kept it, and willpower had very little to do with it.

I have ADHD. My brain has two settings: can’t focus on anything, or hyperfocused on one thing to the exclusion of everything else. There’s no middle gear.

Side projects used to die for me in a very predictable way. I’d get an idea, ride the hyperfocus for a few nights and get the fun part working. Then I’d hit the long middle bit. Auth, settings screens, billing, deployment, the fifteenth form. The wave would break somewhere in there and the repo would go and sit with the other ones on my GitHub.

Agents eat that middle bit. The hyperfocus lasts long enough to reach the end now, because the end is so much closer. I get to stay on the stuff my brain finds interesting, the design calls and the weird problems, while the parts that used to drain me get done and land in front of me for review.

The feedback loop matters too. An ADHD brain wants its reward now, and going from an idea to a running thing in one evening is about the fastest reward software has ever handed out. It’s very hard to lose interest in something when it keeps working.

I haven’t changed my mind on everything. In June I wrote about AI brainrot and I still think it’s a problem. If you hand everything over and stop reading the diffs, you’ll get worse at this job while feeling faster than ever. I deal with that by reading the code that matters, running the tests myself and keeping the decisions mine. The agent writes the code. I own it.

If you’re a developer still holding out the way I was, I get it. It feels like giving up the part of the job you’re proudest of. But the edge you’re protecting is gone. Someone with your experience and an agent will out-ship you, and you may as well be that person.

Pick something you’ve been putting off for a year and give it to an agent in small pieces. Review what comes back like a pull request from a sharp colleague who’s never seen your codebase. Pay for a proper plan so you’re not rationing every prompt. You’ll know within a week whether I’m full of it.

I’ve been writing software for a long time and I can’t remember feeling this optimistic about it. My ideas list is as long as it’s ever been, but for the first time in years it looks like a list I might get through.