Working with AI in 2026

I have used AI coding tools for a long time. The recent Claude Opus 4.6 release felt like a line crossed. Older Claude felt like a smart pair programmer that sometimes wandered off or left rough edges. This round it finishes the task, sweeps up, and hands back something I can run. That shift has me staring at the job market and at what “developer” might mean in a few years.

A friend mentioned that most of the Qur’an rests on roughly five hundred root words: learn those meanings and you cover a large share of the text (often cited around three quarters). I went home thinking a focused app could work.

I needed:

• every word in the Qur’an
• frequency counts
• transliteration
• translation
• audio of each word in Arabic

No single public database had the full set. One site had words and counts, but transliteration only for part of the list. I took it anyway.

I asked Claude to write a scraper. It dumped everything it could into JSON.

Translations were still missing. Another script, same site, pulled Qur’anic glosses per word back into the file.

Transliteration came next. A small script mapped Arabic to Latin letters. The first pass ignored a few rules of Arabic spelling; I pointed that out, and the next version handled them.

Audio meant thousands of clips. I had the model generate one file per lemma, then noticed leading and trailing silence. A trim script stripped the dead air.

At that point I had data. For the app itself I wrote a long spec: screens, state, edge cases, anything I could think of. I expected bugs, ugly UI, or dropped requirements.

Instead it ran on the first try. The home screen felt bare, so I asked for charts and summary stats. It added five I would not have sketched myself. The codebase is large; a strong human engineer might have needed weeks. I had something shippable in about two hours. I read the output. Most of it looks like something a senior dev would sign; a few choices I would refactor, but nothing embarrassing.

The unsettling part is less “it is good” than “it got this good this fast.” 4.5 already felt like a jump; 4.6 arrived almost immediately and cleared that bar again. The public narrative is recursive self-improvement and steeper curves ahead. It does feel like a new default, and not one we can pretend away.

A year ago Claude was a tool in the stack. A year out, I expect it to rival the best programmers alive on raw throughput. Past that, guessing a decade feels silly; the error bars are too wide.