AI-Driven Development Is Faster. So Why Am I So Tired?
AI-Driven Development Is Actually Great
AI made me faster.
I can brainstorm with it, ask it to explain unfamiliar technologies, explore implementation options, and generate a lot of code in a short amount of time.
So when I started using AI more heavily this year, I expected development to become easier.
Instead, after working with other developers on an AI-heavy codebase, I started feeling tired. 😨
So, What’s Making Me So Tired?
It was always the same problem: context. Where it lives, who can see it, and who has to carry it.
The codebase
The codebase was rewritten entirely with AI. It contains decisions that I can no longer safely understand without AI.
For example, I was asked to change the label order on a chart. The chart went A → C → B, clockwise. The legend next to it went A → B → C.
The mismatch came from the design. When we built it, the AI added a comment saying the order was intentionally different. Then the design changed. Now the chart should match the legend.
A simple change. I just did it and opened a PR.
Then the AI reviewer asked: "The comment says this order is intentional. Is the comment wrong, or is your change wrong?"
I didn't even know that comment was there.
The comment saved the decision, but not the reason. And the AI reviewer had no way to know the design had changed.
So now, making manual changes feels risky.
Access permissions
AI is supposed to reduce the frequency of context switching. Ironically, I'm a contractor, so my access is limited. I have to do the context switching myself.
I had to search for where the requirements were, and copy information from Notion into the AI.
I had to export the design image from Figma and give it to AI.
I had to explain decisions from conversations in Slack to the AI.
The planning
I started planning in a draft PR, so the AI could read it. I updated it as more decisions got fixed.
Then it became another access problem. Management can't see the codebase.
So before my plan was even finished, management took it over and moved it to Notion. The draft PR became useless.
The Problem Isn't Writing Code Anymore
At some point, I realized that writing the code wasn't the exhausting part.
Explaining things was. 🤯
When I ask AI to implement something, I have to give it a lot of context:
This part is decided.
This part is still being discussed.
Don't change this yet.
This behavior is temporary.
This design isn't fixed yet.
Management hasn't decided this yet.
Then the PR gets reviewed by another developer, who is also using AI. But that AI doesn't have the context I had. It sees the code as if every decision has already been made.
So it flags something that I intentionally left unresolved.
I explain why. Then I try to make that context visible in the code.
The reviewer says the comment is unnecessary.
So I put it in the PR description.
The reviewer says the PR description becomes the squash commit body, so don't make it too long.
I make it shorter.
Then it gets ignored.
That day, I wrote maybe 10 lines of code. I still closed my laptop feeling completely drained.
So, the bottleneck has moved silently.
It's no longer:
"How do I implement this?"
It's:
"How do I make sure every AI and every developer involved in this change has the same context?"
AI Is Fast. Maybe My Development Process Isn't.
I still want to use AI-driven development.
The productivity gains are real. But I think I am still figuring out what development looks like when multiple AIs are involved in the same codebase.
I tried a draft PR. It ended up in Notion. I still don't know where it should live.
Maybe we need clearer answers to questions like:
Where should the source of truth live?
How do we represent decisions that haven't been made yet?
How do we prevent AI from repeatedly rediscovering the same context?
I don't have good answers yet. But I keep trying to figure it out.
