Slow Code

This week in my personal work, I’ve been writing a lot of “slow code.”

Asking AI lots of questions. Reading docs. Trying to actually understand and internalize the answers. Mostly React, TypeScript, and Bash this week.

It’s been a nice change of pace from the last six months of frantic “Gas Town”-style vibe coding.

With slow code, I know what’s under the hood. I can feel myself becoming a better engineer again.

In the age of AI, I’m pretty sure those skills still matter. Even when LLMs generate the code, understanding the systems helps with reviewing PRs, debugging, and knowing when the AI is confidently wrong. Claude last week literally told me to “skip prod verification and just call it a day”.

There’s nothing wrong with “agentic fever”. Here are things I learned during my six months of trying to keep multiple Claude terminal panes churning out code in parallel.

• I learned what GitHub worktrees are and when they do and don’t work well.
• I learned what kinds of coding work parallelizes well and what doesn’t.
• I learned SvelteKit after Claude suggested it, and really enjoyed how fast it made MVP development.
• I spent a few months in the Neon/Cloudflare/serverless world and started learning about the tradeoffs.

I still want to improve my agentic chops. I want to automate more workflows, build faster, and keep exploring Codex and agents.

But the different modes still matter. Sometimes parallelized chaos is the right call and sometimes it’s slowing down long enough to really understand what’s happening underneath.

Even in the age of “ship ship ship.”