Learn how to define work before implementation, verify generated changes, keep delegated work reviewable and recover when an approach fails.
You should be comfortable reading, changing and testing code. Newer engineers get a method to follow. Experienced engineers examine decisions they may otherwise make by instinct. You do not need to use a particular model or editor.
Small, well-defined tasks with a quick check are where coding assistants can be most useful. Long, ambiguous repository work can hand that gain back through review, rework and regression cleanup.
A controlled experiment found developers completed one defined JavaScript HTTP-server task 55.8% faster with GitHub Copilot.
METR's early-2025 randomised study found experienced open-source developers took 19% longer with AI tools on real issues in repositories they knew, even though they believed they had been faster.
Lee calls the lost time review churn and expansion: plausible changes grow, engineers find regressions and the time saved generating code is spent establishing what is safe to keep.
Read the controlled Copilot study, METR's early-2025 study and METR's 2026 update. The later update says newer tools probably improve results, but selection effects prevented a reliable current estimate.
The examples below come from the current course repositories. They describe the learner task without revealing the facilitator checks.
Treat generated work as untrusted until you have checked it independently.
Define ambiguous work before asking a coding tool to implement it.
Give a coding tool only the work you can still review properly.
Change course when the evidence shows that the current approach is wrong.
Lee explains the rule and the engineering risk behind it.
You inspect the task, use your coding assistant and run your own checks.
You explain what you accepted, changed or rejected and show the evidence.
The group compares approaches and identifies what to carry into normal work.
The current exercise library spans all four languages. Python covers the complete course; selected modules also have TypeScript, C# and Rust variants. Lee confirms the exercise set before booking.
Each repository includes a task, existing code and visible tests. You keep the final change and written work you produce.
Feedback covers the result, the checks you ran, the scope you controlled and how you responded to new evidence.
Model choice, prompt style and token use are not scored. A single session gives limited evidence and is not presented as a certification.
The course is live and unrecorded. You control what you share with Lee and the group.