I Spent 20 Years Building Web Apps. Then AI Taught Me What I Missed

I have been building web apps for about 20 years. PHP when it was uncool, then when it was cool, then when it was uncool again. I have shipped for clients, for employers, and now for my own product. So when AI coding tools showed up, I expected them to save me typing. Faster boilerplate, faster lookups, a smarter autocomplete. That is not what happened. What happened was more uncomfortable, and more useful.

For two decades the spec lived in my head. I would read a ticket, form a picture, and start typing. It worked because the only person who needed the picture was me. An AI has no access to my head. If I give it a vague instruction, I get vague code back, every single time, and the model is not to blame. The first weeks I thought the tool was weak. Then I noticed that when I wrote down exactly what I wanted, inputs, outputs, edge cases, what should happen when things fail, the results jumped. The tool had not improved. My thinking had been forced out into the open, where its holes were visible. Twenty years, and I had skipped the skill of specifying, because I could always compensate with familiarity.

I believed my 20 years were 20 years of accumulated judgment. Some of it is. Knowing that the hard part of a feature is the migration, not the code. Knowing which shortcuts explode in month six. That judgment is real and AI made it more valuable, because now I apply it to ten times more code than I could write myself. But a lot of what I called experience turned out to be habit. The way I structured controllers, the way I named things, patterns I defended because I had always done them. When a model suggests a different approach and I ask myself why mine is better, sometimes I have a reason. Sometimes the honest answer is that I have just been doing it that way since 2009. Those two feel identical from the inside. They are not.

The skill I lean on most now is one I undertrained for years, reading code I did not write, fast, with suspicion. Generated code looks confident. It compiles, it runs, it demos. Whether it is correct in the corners is a different question, and answering that question is now most of the work. The developers who will struggle are not the older ones, they are the ones of any age who cannot review. Writing code is becoming cheap. Judging code is becoming the job.

Write things down before you build them, even when nobody asks. Learn to explain your decisions, because one day something will challenge them, and if the only defense is seniority, you have no defense. And do not confuse fluency with understanding. I typed fluently for 20 years. Understanding is what survives when the typing gets automated.

I am not worried about AI taking the job of a person who thinks clearly. I am watching it quietly absorb the parts of my job that never required thinking, and honestly, it can have them. What would an honest audit of your own 20 years, or 5, or 2, leave standing?