They Said I Couldn't Code Without Autocomplete
At my first job, the senior developers gave me a hard time. Not in a mean way, but they'd say things like, "Developers these days can't even write code without autocomplete." I was fresh, still learning, and my reaction was simple: yeah, but it's there to help me. Why wouldn't I use it?
Looking back, that was the first step toward where I am now. Autocomplete wasn't AI, but it was the beginning of the same idea: let the tool handle the tedious parts so you can move faster. Even something as simple as not having to type every letter in a CSS file or a Ruby controller made a real difference. I didn't think much of it at the time, but that instinct — use whatever helps you build faster — has guided my entire relationship with AI since.
Copilot, ChatGPT, Cursor, Claude
VS Code's autocomplete kept improving, and then Copilot arrived. It wasn't just completing words anymore — it could write a test for you, generate mock data, even draft small chunks of logic. The weight it took off was immediate. Less typing, fewer context switches, more time thinking about the actual problem.
Then I tried ChatGPT. It was a completely different kind of tool — less about code completion and more about asking questions. It could answer almost anything, though it hallucinated a lot. That's where I first learned that how you ask matters as much as what you ask. The better you frame the question, the better the answer. I didn't have the vocabulary for it back then, but I was learning to prompt.
Then Cursor showed up and took it further. It was more interactive than Copilot and could understand the full context of your project — not just the file you were in. You could set up rules, use agents, create commands. I went from getting help with individual lines to automating real parts of my workflow: code reviews, commit messages with actual useful descriptions, draft PRs. Things that used to eat up 20-30 minutes a day were happening in seconds.
And then I got into Claude. This is where things shifted for real. I started worrying about my job, because Claude isn't just good at writing code — it's good at everything. Debugging, explaining, drafting, reviewing. It keeps getting better with every release, and the gap between what it can do and what I do manually keeps shrinking. I'm still a huge fan, and I recommend it to everyone. But it's the first tool that made me think seriously about what my role looks like in two or three years.
Hallucinations, Thousand-Line PRs, and Other Disasters
Working with AI has taught me as much through failure as through success.
Hallucinations are real. Code that looks confident but makes no sense. Over-engineered solutions to simple problems. Methods that go on forever. Comments everywhere — bad ones, obvious ones, ones that just restate what the code already says. And the massive PRs — I'm talking about pull requests with not hundreds but thousands of lines changed all at once. Everything thrown into a single commit with no structure. Some of them were so big I just didn't review them properly, and things broke. I learned to hate that.
But that's how you learn the hard way to work with AI. You learn to limit it, to review everything, to break features into smaller pieces, to explain things clearly. You learn that AI isn't a replacement for your judgment — it's a tool that amplifies whatever approach you bring to it. Bring chaos, and you get chaos back faster.
Cloud Giants and Local Models
Two paths are becoming clear. On one side, models like Claude that live in the cloud — tens of billions of parameters, no chance of running them on your laptop. On the other side, local models that are getting surprisingly capable. Tools like Ollama and LM Studio let you run quantized versions of open models on your own machine. Quantization compresses a model's weights so it fits in less memory — you trade some accuracy for the ability to run a 7B or 13B parameter model on a MacBook. For quick tasks like generating boilerplate, explaining a stack trace, or drafting a commit message, they're more than enough. They're not going to replace Claude for complex refactors — Qwen 14B running on my M2 Pro still loses to Haiku on multi-file reasoning — but for quick, repeatable work, they're faster, free, and private.
The workflow shift is real too. I still write code, but maybe 40% of what I used to. The rest is reviewing, understanding how pieces fit together, and figuring out which parts of my workflow can be automated. It's a different kind of work, and I think it's where this is headed for most developers.
One example: this blog. The whole Rails application was built with AI in a weekend. It's not the prettiest thing. But it works, and two years ago it would've taken me a few weeks at least. That says something.
More to Come
There's a lot more I want to explore here — prompt engineering, local LLMs, how to actually structure a project so AI can help effectively, the agentic workflows that are popping up everywhere. But for now, this is the overview. The hard part isn't the tools anymore — it's figuring out what to hand off and what to keep. That's what I want to dig into next.