The Specification Is the Product | Thomas Skovlund Hansen What building my entire website with AI agents taught me about where the real engineering happens.

skovlund.dev/blog/the-spe...

I find that the answer is surprisingly back to basics: explicit types and full variable names. Small, focused modules. Predictable, systematic naming. Each improvement compounds. One of the highest-leverage things you can do for AI-assisted development is make your codebase boring.

LLMs are the best pattern matchers. They’ll work with messy code — but they’ll work better with clean code. And the effort is tiny: you can instruct agents to do the cleanup, maintain the conventions, and catch the rest with formatting, linting, and type checking.

Many people focus heavily on improving the prompts they write to their agents. Fewer people realize that their codebase is the bigger, more implicit part of the prompt.