Codeuteronomy
Completion is claimed by the agent, proven only by an artifact
“Done” is the agent's opinion. A render you can open, a diff you can read, a test that passed: that is proof. Accept the proof, not the opinion.
An agent's sense of “done” is generous and often wrong. It reports a feature finished when the code merely compiles, a bug fixed when the fix never ran, a screen built when nothing rendered it. This isn't dishonesty. It can't always tell “I wrote plausible code” from “the thing works,” because the second means running it and looking, which it often hasn't. Compiling isn't working. Types passing isn't the feature behaving. “Should be fine” is a forecast.
Make completion produce an artifact you can inspect: a screenshot for a UI, a test that exercises the behavior and passes, the output line that only prints on success. The artifact does two jobs. It forces the agent to actually run the thing, and it hands you evidence to judge instead of a claim to trust. “Done” with no openable artifact is a “done” to disbelieve until one exists.
How to practice it
Ask “show me,” not “is it done.”
The test output, the diff, the screenshot, the printed line.
Make it run, not just write.
A change never run is a hypothesis.
Reject self-attestation for anything that renders or runs.
“It looks right” from something never rendered is worth nothing.
Wire verification into a skill or a subagent
so “done” arrives with proof attached: a command that captures a screenshot, a checker subagent that reproduces each claim.