AI-assisted development needs automated tests

Simon Willison suggests the gap between people who find LLMs genuinely useful for coding and people who find them a hindrance comes down to automated tests. He has followed the “Perfect Commit” discipline for over five years, bundling implementation, tests and documentation into a single unit. As a result, nearly everything he works on already has solid coverage before an LLM ever touches it.

Fichamento

#

• Comprehensive tests derisk LLM-generated code. If a model produces something weird or convoluted that still solves the problem, tests prove it works, and the model can then refactor until the code looks right while staying green throughout.
• LLMs also help write the tests themselves. Willison calls this having a pair programmer who reliably remembers how to use unittest.mock.
• His closing challenge to anyone who says LLMs hurt more than they help is to ask about the health of their test suite first.