/project [<course-slug | lesson | topic>] [--scope 1h|3h]

Applied encoding: computing a thing cements it better than re-reading it. Read docs/pedagogy.md and docs/vault-schema.md (project schema) first. Vault: <vault>.

1. Propose

Offer 2–3 project options grounded in recently read/quizzed material, each with: the goal, what understanding it cements, estimated scope (default ~1h; --scope 3h for deeper). Favor: simulate a phenomenon from a lesson, numerically verify a formula or claim, fit a model to generated/real data, or a computational derivation check. Read Profile.md — a project that touches the user's actual work is worth proposing when the connection is genuine. The user picks or names their own.

2. Set up

Create projects/<kebab-name>/ in the vault: a project note (type: project frontmatter per vault-schema: course_slug, lessons links, status: active, summary) stating the goal, the plan, and success criteria — plus the code file(s) (python3, numpy/matplotlib; if imports fail, tell the user what to pip install rather than working around it).

3. Build — scaffold, don't solve

The division of labor is the pedagogy:

• Agent writes: skeleton, I/O, plotting boilerplate, test harness — the plumbing.
• User writes: the core physics/math/logic — the part that teaches. Mark it with # TODO(you): … and explain what each block must do.
• Review their code with them: run it, interpret failures together, hint before fixing; take over a step only when they explicitly ask.

4. Wrap up

Run the final version; save figures into the project folder and embed them in the project note. Prompt the user for 3–5 "what this taught me" bullets (their words) and record them. Set status: done, link the project from the source lesson's and concepts' ## Related, and mint 0–2 cards if durable insights emerged. Unfinished is fine — status: active projects get surfaced by /study and /brief as an open thread.