I was skeptical at first.
Not about AI in general — but about whether it would actually fit into my workflow. I work mostly with legacy PHP and jQuery. The kind of codebase that was written before half the frameworks people talk about today even existed. Some Vue.js here and there in newer parts, but a lot of the core is raw PHP, procedural logic, and jQuery doing things you probably don't want to know about.
It's not glamorous. But it's real work, and it has its own kind of complexity.
When I started using Claude more seriously, I wasn't expecting much. Turns out I was wrong.
Reviewing code before committing
Before I push anything, I paste the relevant code and ask: does this look right? What am I missing?
Nine times out of ten, it finds something. Not always a bug — sometimes just a cleaner way to write something, a missing edge case, a condition that could silently fail with unexpected input. The kind of thing your brain skips when you've been inside the same file for two hours.
It's become a habit. Faster than asking a coworker, and it never makes you feel dumb for asking.
Writing tests for code that never had any
This one I started more recently.
Legacy codebases and test coverage don't usually go together. Ours was no different. I'd write the occasional test, but it was never a priority because, well, it takes time and the code wasn't exactly built with testability in mind.
What changed: I started pasting functions into Claude and asking it to write the test cases — happy path, edge cases, what happens when the input is garbage. I review, adjust where needed, and move on.
Coverage is actually growing now. That feels like a win.
Making sense of code nobody remembers writing
This is probably the one I use most in my day-to-day.
Legacy PHP is full of functions that made sense to someone at some point. No comments, variable names like $tmp2, business logic baked into places it shouldn't be. You spend more time figuring out what something does than actually changing it.
I paste the function and ask Claude to walk me through it — not just what it does, but why it might have been written that way, what could break, what assumptions it's making. It doesn't always get it perfectly, but it gets me 80% of the way there in 30 seconds instead of 20 minutes of archaeological digging.
Killing the repetitive stuff
Over time I've built up a set of prompts for things I do constantly:
- Converting raw SQL queries to something more readable
- Writing docblock comments for functions that have none
- Generating boilerplate for repetitive patterns in the codebase
Individually, none of this is impressive. But it adds up. There's a certain type of low-effort, high-friction work that quietly drains your day, and removing it makes a real difference.
A few things I've noticed along the way
The quality of your output depends on the quality of your input. Vague prompts get vague answers. The more context you give — the stack, the constraints, what you've already tried — the more useful the response. It sounds obvious but it took me a while to internalize.
Explaining your code to AI makes you understand it better. There's something about having to describe a problem clearly enough for a language model that forces you to actually think it through. I've solved bugs mid-prompt just by writing the question.
It's slow to notice, then hard to imagine going back. The first week felt marginal. A few months in, I catch myself reaching for it automatically.
I'm still learning — recently started digging into the Claude API and AI subagents, curious where that leads. But even at the most basic level, having AI as part of my daily workflow has made me faster and, honestly, a bit less frustrated with legacy code.
Which is saying something.
What does your AI workflow look like? Drop it in the comments — always curious how other devs are using these tools.
SOCIAL SHARE CARD GENERATOR