Test

A test is a script that refuses to let last night's bug happen twice.

You do not re-click the whole app by hand after every change. You write a small program whose only job is to run another piece of code and confirm it did the right thing. Give it this input. Expect that output. Green means it still does. Red means you just broke it — including future you, including the agent that "refactored" at 3 a.m.

A unit test pokes one function. An integration test pokes a few pieces wired together. Neither is a judge clicking through the demo. That's a different, slower check. Start with the one path that already bit you.

Why this matters when AI writes the code

An agent can produce a lot of code fast. It can break a room three files away just as fast. Tests are how you keep it honest. The agent makes a change, you — or CI — run the suite, and if it snapped something you cover, you find out before you deploy. An eval is a test for prompts. A type check is a test the compiler already wrote. Pair them.

Green does not mean nothing is broken. Green means the things you pointed at still work. That's why a real net is a lot of good tests, not one that asserts true.

The CLI already has the verb: bun test. The builder can scaffold a project that knows it. CI is the teammate who runs it on every pull request so tired you cannot skip it.

At a hackathon, write the tripwire for the one query the demo dies without. Don't write a framework. If the agent lies, the tripwire screams before the judge does.

What this unlocks

Changing code stops being a held breath. You make the edit, you run the suite, and "I think it works" becomes "I checked the parts I care about." Agents stay fast without flying blind. The bug that bit you once has a name and a red light, and it does not get to bite the judge.