Session 1. Configure the assistant and the repository instructions — Mon 14 Sep

Conclusion

You configured an assistant that reads AGENTS.md before it edits, and you read all three policy files to see why one canonical file beats three copies.

What you did

• You wrote the acceptance checks into the file itself: uv run bootcamp check ch01 and uv run ruff check . decide done, and the assistant never claims a command ran unless it ran.
• You ran the loop on a tags filter for search_documents: plan, restrict, smallest diff, test, review.
• You rejected one change and wrote down why. That record is the session's artifact.

The failure you handled

An unauthorized file change. You asked for an edit outside the allowed files, and either the assistant refused or you rejected the diff. Silence was the only fail.

What to carry forward

Day one also set the mental model: request → context → model → tool → verification. You own everything but the model.

The builder-kit under builder-kit/ is what repository instructions look like when they travel; session 10 has you write a skill of your own.

Into session 2

Session 2 assumes the preflight cell is green, FakeLLM.complete(system, user) makes sense to you, and your assistant works the TODO(you) lines while you run the check cell.

From tomorrow, every exercise ends in check(...) and every notebook ends in review("chNN"). You read the verdict, not the assistant.

Homework

The ambiguous instruction gave you your homework: the sentence that would have made the assistant ask, added to AGENTS.md.