Read AI output like compiler output

Most of us never read compiler output. We read what the compiler tells us: what changed, what broke, where to look. I’ve started treating AI-generated work the same way.

When an agent hands me a diff, a PR or a long answer, I don’t read it top to bottom. I read three things: the interface (what changed), the errors (what can break) and the hotspots (where I should look or act).

The analogy has a limit. A compiler is deterministic and heavily tested. An agent isn’t. The contract is also self-reported, so I still check the hotspots myself and rely on tests and CI.

Ask for a contract

The fastest way to get those three things is to ask for them:

Give me a contract for this change.
  Problem: one sentence
  Change: one sentence
  Risk: one sentence
  Rollback: one sentence
  What I must do: one sentence

If the agent can’t fill a line clearly, that’s the part I go and read myself.

Catching up

The same idea works after a day away:

List all PRs and issues since yesterday.
Give a one-line outcome for each.
Mark only the ones that need my decision.

Everything not marked waits until it becomes a blocker.