Robert Ta explains how to pair human engineers with AI agents: humans write exact specifications and test criteria, AI generates code and fixes it until tests pass. The leverage isn't in the AI writing code, but in humans getting good at judging quality and writing clear specs.
The best and fastest AI engineering teams in Silicon Valley will have repos with literally zero handwritten code. The humans focus on the English and the specification. The agents write the code and implement, but the humans use their brains to figure out what is good and what is bad. That's Spectra of Involvement. Here's how it works with Cloud Code. Step 1: Write what you want in plain English. Be precise. For example, here's a feature: password reset via email. User clicks "Forget password." The system will send a reset link valid for for 24 hours. The link then leads to a form requiring an 8+ character password with variation. Can't be an old password. Expired links then show an error with an option to request a new one. Every sentence there is a testable condition. Determinism. No ambiguity. Step 2: Write the tests before the code. This isn't a new concept. Test-driven development has been a thing in the industry forever, and it's even more important in the age of AI because it's non-deterministic, these tools. So you gotta give it clear exit criteria. Step 3: tell Cloud Code, implement the spec, all tests must pass, then code review it and bring me the results. Cloud Code then will read the spec that you've defined with what good is and the exit criteria so it knows when it's done. It'll write the implementation, it runs the test, it reads the errors, it fixes the code, run the tests again, and it's in that loop just like a good engineer would because it has a good spec. You only get high leverage though when you put your brain on what is good, what is bad, and what's the exit criteria.
