The two-day outline we settled on after running it six times, and the parts we removed because they did not survive contact with real repositories.
Six cohorts in, the outline for the Claude Code course is shorter than the first version by about a third. What was removed is more interesting than what stayed.
Walking through every capability takes an hour and produces notes nobody uses by Thursday. We now start with one small task in the participant's own repository and let the questions pull the features in.
For code work, the useful habits are concrete: provide the file, state the acceptance condition, and ask for a change small enough to review. That is a fifteen-minute conversation, not a module.
The session that changes behaviour most is about what the agent may not touch. Secrets, production systems, migrations without review, and destructive commands. Teams that skip this have an incident within a month.
Time per merged change, compared with the month before. It is imperfect, but it survives an argument and it is hard to fool.