AI Is Forcing Us To Write Good Code - by Steve Krenzel

The team discovered that demanding 100% test coverage is the single most effective guardrail when letting LLMs write code. At full coverage the model can no longer hide behind untested lines; every change must be proved with an executable test, turning ambiguity into a concrete to-do list. This phase change makes the agent's work predictable and dramatically reduces the risk of hidden bugs.

File organization and naming become a second interface for the model. By giving each module a clear, purpose-driven path and keeping files small, the LLM can load full contexts without truncation, improving its reasoning and speeding up iteration. The same principle applies to the test suite: the team engineered a fast, highly concurrent test runner that spins up isolated databases in seconds, allowing the agent to run the full 10,000-assertion suite repeatedly without slowing down.

Automation is taken to the extreme. Git worktrees, environment isolation, and cheap, repeatable linting and formatting hooks keep the agent on a short leash, forcing small, verifiable changes. Docker or similar isolation ensures many worktrees can run concurrently without port or state conflicts, letting engineers spin up fresh environments in one or two seconds.

Finally, the codebase leans heavily on a strong type system. TypeScript provides compile-time guarantees that eliminate entire classes of illegal states, while well-named types like UserId or SignedWebhookPayload give the model instant semantic clues. Combined with typed OpenAPI clients and typed database access, the environment becomes a self-documenting contract that an LLM can navigate with confidence.