published on: Sunday, February 1st, 2026
Sidenote: I made an app - completely using Claude Code. Zero lines of handwritten code - It’s a crosswords app which I built mainly for myself as I enjoy solving crosswords. Here it is: Synonymist check it out!
Last year, around May, I was writing blog posts that were titled: You can’t “just use AI”. Now, in January 2026, the picture has changed significantly. I think we are at the start of something insane in software development.
I still stand by the blog post I wrote at the time, because it accurately reflects my thoughts and thinking at that time. It’s kind of still useful as it serves as a marker for LLM capabilities around May 2025 and it’s marvelous how far we have come in just 1 year.
I was first skeptical about claude code and had a bit of an AI-model burnout as new models were being released pretty rapidly all through 2025. I got access to quite a few of the cutting edge models through work, so I could evaluate them as a practicing, professional software engineer - and use them on real tasks.
The models were pretty useful and boosted my productivity - especially when I had to do one-off tasks, or when I needed to codify something which I didn’t have time to do before. But, I still found them pretty limiting. After all, I, a human who goes to meetings, and can have hallway conversations with product managers, and designers, had a lot more context about the problems I was required to solve than any AI model.
Enter Claude Code - powered by Claude Opus 4.5 (I will call it CC from now on)
In December 2025/January 2026 - everything changed. I would no longer write software by hand again unless in very specific circumstances. I got access to CC through work sometime in December. I usually take my annual vacation during December, so I didn’t really spend much time playing with it as soon as I was given access. Once I was back from my vacation, I dug into the new and shiny tool I was given access to.
Before, I started using this model, I had heard anecdotes from other developers, about how amazing this model was. The first blog I read about this model was from antirez(creator of Redis) called Don’t fall into the Anti-AI hype
So, right away, I had social proof that motivated me to look at CC in a different way. When I got a project where I could actually exercise its capabilities, I set up a personal goal: I would write exactly 0 lines of code by hand to implement the product requirement. The new project should be written 100% by CC, with me prompting and verifying its output.
I set this rather ambitious goal, on a pretty hard-to-do project (PDF generation of reports we show in the application with all its related warts and thorns and gotchas) and was expecting it to be somewhat successful. What I instead found was that the model was doing 90% of the work and I had to treat it more like a junior engineer, than like a dumb LLM or a coding agent. It’s no longer a fancy auto-complete. The model exhibits emergent behavior. I don’t know how Anthropic did it, but they have created millions of software engineers on demand.
The amount of leverage I now possess has allowed me to automate large parts of my job completely - and I keep finding new ways to automate the remaining tasks I’m still doing. The way I approach problems now has also undergone a dramatic shift - I no longer care about the code (that was hard to type out) - for all intents and purposes, the high level languages I used to use (TypeScript, Python etc) are now like assembly languages - and I mainly program in english.
This does not mean that I don’t look at diffs, or review code, or I don’t care about the application architecture - rather, it means, the drudgery of typing and having to remember a bunch of design errors in the language I’m using (null versus undefined in JS for instance), is a thing of the past. Programming is, to me, a way to get the computer to do whatever I want. And now I can just tell it what I want - even on large, brown-field codebases with all their ugliness. (Granted, I was involved with this codebase since the start (~2 years ago), and have had a heavy, almost draconian hand in shaping it - for example, the eslint rules are super-strict and we do not allow the use of any in typescript - things like that.. which makes it easy to trust the code generated by Opus)
The most direct way in which my workflow when writing software has changed is the following: I used to put up with a bunch of crap when writing code - things like slow user interfaces on tools I was using (gitlab, jira) etc.. now, I just have CC write a bunch of shell scripts for me which use the respective tools’ apis to accomplish what I want. So I can stay in my terminal without having to context switch. This is how I was able to create a list of stacked pull requests and update the associated jira tickets by literally typing:
Can you look at the git log and identify the tickets? go ahead and update them with the
relevant pull request links by leaving the PR link as a comment on the jira tickets
This is power which Prometheus stole from the gods :)
Most people are still fighting against these tools. I’m certain the war - is already won by the LLMs. They are just way too powerful and capable, and programmers who refuse to learn/use them will be left behind in dust. Whether this leads to job losses is hard to say, but I think it perhaps will. At least in certain sectors where it was previously considered un-economical to automate things via deterministic programming. Now that the cost of building(implementation) has dropped to zero, and the productive capacity of a single engineer has been multiplied 10x (100x?), a lot more stuff is going to get built for sure.
The way I use CC is very different - it doesn’t bear any resemblance to how I was using AI last year. That’s actually pretty nuts. It’s also further evidence of how quickly things are changing in the field - yesterday’s strategies are outdated today.