Lately I've been struggling with a pattern in my work.
AI has made me dramatically faster at generating code. In my office-related work, where I already understand the domain, the stack, and the constraints, my agentic workflow often feels like a 10x boost. Sometimes even more. I can move faster, make better decisions, and still stay grounded in what the code is doing.
But in my personal projects, the story has been very different.
A lot of those projects live in stacks I don't know deeply. And there, the same workflow that feels magical at work starts to fall apart. I'm not getting 10x results. Some days I'm not even matching the timeline I would have hit if I had just written the code manually.
That was the first moment I realized there was a real problem.
The issue isn't just whether AI can generate code quickly. It obviously can. The issue is that the speed of code generation is now faster than my speed of understanding.
When I used to write code by hand, it took longer, but the slowness helped. Writing was also how I built a mental model. I knew what the code was doing because I had to think through it line by line.
AI breaks that loop.
Now I can get working code before I've really earned an understanding of it. And because my mind is already on the next feature, I often move on too quickly. I tell myself I'll come back and study it later. Usually I don't. Or not deeply enough.
That gap compounds.
Feature 1 works. Then feature 2. Then feature 3. By the time I'm at feature 10, feature 1 starts breaking. I go back to fix it, and in the process I break feature 5. Then fixing feature 5 affects feature 8. What looked like fast progress turns into slow, expensive stabilization.
That's the part I keep running into: AI compresses implementation time, but not comprehension time.
I can absolutely build features faster than before. But getting them production-ready still takes just as long, sometimes longer. The early phase got cheaper. The later phase got heavier. I'm not paying less for the work. I'm just paying later, in confusion, regressions, and rework.
What I'm starting to believe is that AI mostly amplifies the understanding I already have. In areas where I'm strong, it feels like leverage. In areas where I'm weak, it can create the illusion of leverage while actually pushing me further away from the code.
So maybe the real bottleneck in AI-assisted development is no longer code generation. It's comprehension.
AI has made code cheap to produce. It has not made code cheap to understand.
And if understanding falls behind, the time saved upfront comes back later with interest.
What I'm changing
Over the next few months, I want to work differently:
- I'm going to stop treating generated code as finished work. If I can't explain it clearly, it isn't done.
- I'm going to be much more careful using AI in stacks where I don't have deep intuition already.
- I'm going to ship smaller batches so the comprehension gap doesn't compound.
- I'm going to ask AI for smaller pieces and better explanations, not just bigger code dumps.
- I'm going to treat production-readiness as its own phase, with time set aside for understanding, cleanup, and edge cases.
That's the shift I'm trying to make.
Not using AI less. Just using it in a way that keeps me close to the code, so I can still trust what I'm building a week later.