An agent with filesystem access is fast, confident, and wrong often enough to matter — and a research repo is the most expensive place for that combination, because a plausible edit to an analysis script produces plausible numbers. The skill being assessed is not prompting. It is bounding the task and reading the diff, which are the two things that make the speed safe.
Claude Code documentation. It covers operation — what the tool can do, how to run it. It does not rule on what a research repo should let an agent touch, which is a question the standard answers and the documentation cannot.
Whatever governs the files it touches. Parts 1–2 for src/, Part 3 for notebooks/,
and Part 4 rules on the boundary between them. An agent does not get an exemption from the
Part that governs the directory it is editing.
The honest outcome of this exercise is sometimes “it was faster to do it myself.” Report that if it happens — it is a finding about the boundary of the tool, and it is worth more than a success.
Transcript excerpt plus the diff you accepted, in that week’s WA entry.
This replaces your arithmetic entry for the week you do it. Submit it to that week’s WA assignment in Canvas — the same one you would have submitted anyway. It is not an extra piece of work on top of the log.
You do not submit anything to the milestone yet. When all four skills in M2 are done, submit the M2 — Reproducible Code assignment, which is just an index saying which week each one went in.