Lectures

Tommy comes in as a champion in the morning, only to find Tony’s portal in ruins: /rewind is dead, and the rollback history was wiped out along with the night session. Panic gives way to the practical question, and what matters now is figuring out exactly what happened. The night log, written to a file on disk, saves the day.

The strong model’s limit was burned through overnight — Alex is silent. Tommy checks /usage, moves Alex to a weak model through /model, reads the night log step by step, and admits three of his own mistakes: “allow everything,” “don’t spare the limits,” and a huge task for the whole night with not a single brake.

The red error wall is not a mess but a crime scene map. Alex teaches Tommy how to read a stack trace (type, file, line), tell apart “where it blew up” and “where the mine was planted,” and connect two clues — the stack trace and the overnight log. Rule: facts first, agent second.

The rule of the day: done is not when the agent says so, but when you prove it by checking. Three independent eyes — tests, build, live run. The agent can lie even about the check, so we look at the real output. A precise clue via @, an incident card for one symptom.

Tommy fixes from an incident card, not by vibes: a narrow frame, one symptom, and a success metric defined in advance. He demands real output from the agent instead of “done,” supplies a precise clue through @, proves it with a three-part chain of evidence, and leaves an EVIDENCE_LOG. New rule: no proof, no result.

Two level articles in one list: how a test works and what pass, fail and skipped mean in the report, and how the model differs from the harness around it in Claude Code.