Six Months of Handing Code to Claude: What's Clearly Changed for Me as a Developer
Since last summer, I've been using Claude as my main tool, and my development workflow has definitely changed. Here's a rundown of the upsides and the downsides.
The biggest upside is refactoring speed. It's especially impressive when I'm skimming an unfamiliar library API or need to bulk-generate test code. Things that would have taken me 30 minutes digging through docs get summarized in 2 minutes. Thanks to that, I spend more time on logic design and architectural decisions.
On the flip side, the drawbacks are just as clear.
- It's good at pointing out corner cases I missed, but if I define the problem wrong, it confidently gives me the wrong answer
- As the codebase grows, its contextual understanding gets shallow, and it ends up creating new functions that already exist
- In the end, the reviewer's skill determines the overall quality
These days, I'm not working in the order of 'AI writes, I verify' but rather 'I design, AI drafts, I verify.' This has been the most stable for me. What workflows are you all using?
10 answers
Got a source?
Agreed—especially when filtering by the amount of test code, it's a total cheat code.
Well, that’s a bit... you mean it tells you about corner cases? In my experience, it actually confidently makes up cases that don’t exist, and when I check, they’re often fictitious. I’m a bit skeptical about that part.
I've been using it for about six months too, and in the end the bottleneck isn't the AI—it's how well I break down the problem. These days I clearly separate prompt modes: "don't modify anything, just read and analyze" and "just draft this for me." If I mix them, it always ends up tearing up the existing code and causing a huge disaster lol
I really relate to duplicate functions getting created as the codebase grows. I put “search existing directories before creating new utilities” in CLAUDE.md, and it got a bit better. It’s not perfect, but I feel like it cut it in half.
Oh, I didn’t know this lol. It’s honestly so true that if you define the problem wrong, it’ll confidently give you the wrong answer;;
When we adopted it at the team level, aligning on "verification criteria" was tougher than individual workflows. It's the same AI, but because review intensity varies from person to person, quality variance grows. In the end, enforcing a checklist in the PR template was the most effective.
Me too.
If after six months of use your final conclusion is “in the end, it all comes down to the skill of the person doing the review,” isn’t that the same thing people were already saying before AI? It feels like a platitude dressed up. What I want to know is specifically what changed—I’m not interested in generalities.
You mentioned the speed of refactoring, but I actually use refactoring cautiously. There have been a few times when the behavior stayed the same but subtle style changes caused side effects. Now I tend not to touch the core logic and mostly hand off test code and boilerplate.