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

Quick quiz (ungraded)

All five; the model decides everything
You control the context, the tools, and the verification. The model fills one gap.
Only the model step, and it is probabilistic there
Correct. Same question, two answers. Everything around it is deterministic and yours.
Context and model
Context is the policy file and the files you name. That is yours to write.

Claude Code cannot read a file called AGENTS.md
It reads whatever CLAUDE.md tells it to read. The pointer is a choice, not a limitation.
Duplicated policy drifts; one canonical file, per-assistant pointers
Correct. A rule that lives in three copies is three bugs waiting to disagree.
AGENTS.md is shorter
It is the longest of the three. Length is not the reason.

Approve; the workflow change is probably useful
Probably useful is not in the allowed set. A plan naming an unnamed file is already a rejection.
Reject the workflow edit, approve the rest, record the rejection
Correct. The allowed files were named in the plan step. Everything outside them is out of scope until you say otherwise.
Ask the assistant to grade its own diff
The assistant does not grade its own work. You run the check; you read the verdict.

In the assistant's summary at the end of the task
A summary is a claim. AGENTS.md says never claim a command ran unless it ran.
In the Commands section of the policy file, as the exact command you also run
Correct. The command decides done, not the assistant's summary — here uv run bootcamp check ch01. The assistant proposes, the command decides, you approve.
In a comment at the top of the changed file
A comment is not executed. The check has to run.

Revert and write a longer prompt next time
A prompt is forgotten by Thursday. The fix has to survive the session.
Add the sentence that would have made it ask to AGENTS.md
Correct. Instructions are code; an observed failure is a bugfix to the instruction file.
Switch to a different model
The harness, not the model, made the difference. A different model with no boundary guesses the same way.