Hey, this is my first post here. I am a backend and cloud developer, but I also use AI to build frontends. I have been thinking a lot about where AI coding is going. I wanted to share something that has been bugging me and see if others feel the same way.
The future we all want
Imagine this. You describe what you want, AI writes the code, you ship it to production. Done.
That is the promise. And honestly, we are getting close. Tools like Cursor, Claude, and Copilot can generate a full React component in 30 seconds. A hook, an API route, a page layout. It is wild how fast the generation part has become.
So if AI can write code this fast, why are we not shipping faster?
What actually stops us
Here is what I noticed after tracking my own workflow for about a month.
The writing part is fast. The shipping part is not. And the bottleneck is not what I expected.
It is the review.
Every time AI generates code, I need to read through it line by line. Does this component actually do what I asked? Did it handle the edge cases? Is it following the patterns we use in this project? That takes 8 to 12 minutes for a single component. Every time.
And here is the part that really got me. When I run the same prompt again, I get different code. Not wrong code. Just different. Different variable names, different structure, different ways of handling the same thing. Both versions work fine. But they look nothing alike.
So I can never say "oh I have seen this pattern before, it is fine." Every generation is a fresh read. The review never gets faster.
I started calling this the review tax
You generate code fast. Then you spend 10x longer making sure it is correct. And that 10x cost shows up on every single generation. The more code AI writes for you, the more time you spend reviewing.
The speed advantage starts to disappear when you zoom out and look at total time.
Why this happens
The root cause is simple. AI output is not deterministic. Same input, different output, every time.
With a compiler, you change a line and you know exactly what changes in the output. You build intuition over time. With AI code generation, there is no intuition to build. The output shifts every run.
This is fine for creative writing. You want variety there. But for code that needs to go to production? You want the opposite. You want predictability.
What would actually get us to that future
If AI generation were deterministic, the whole game changes.
Same input, same output. Every time. You review a pattern once, approve it, and then trust every future instance of that pattern. Your review time stops growing with the amount of code generated.
That is when AI coding tools actually deliver on the promise. Not just fast generation, but fast shipping.
What we are building
This is exactly what we are working on at promptend.ai. A coding agent for frontend developers where the output is deterministic. We are still early, still pre-launch. But the idea is straightforward. Cut the review tax so developers can actually ship AI generated code to production with confidence.
Honest question
I am curious. How much of your AI coding time is spent reviewing versus prompting?
My guess is most of us have not measured it. I had not until recently. The numbers surprised me.
Would love to hear from others, especially frontend devs working with React and Next.js. Is the review tax real for you too, or am I the only one noticing this?
Top comments (0)