You can drive it.
Now learn to engineer with it.

Ten modules. Four exercises.
About 13,000 words of it.

1. Setup, done rightThe environment, the permission model, the project structure — the parts the quickstarts skip for speed.
2. Memory, the deep diveHot files vs cold files, provenance weighting, pruning discipline, and the exact failure mode of a memory store nobody maintains.
3. CLAUDE.md as an operating systemThe instruction hierarchy, token budget as a real constraint, and how to write an instruction file that still works at 300 files.
4. Hooks & automationThe harness-level events that fire automation the model can't forget to run — because the model never runs them. The harness does.
5. Building skillsPackaging a capability so you invoke it by name instead of re-explaining it. Frontmatter, scoping, and when a skill is the wrong tool.
6. Multi-model routingStakes × verifiability as the actual decision function. Cheap models for grunt work, expensive models for judgment, and how to tell which is which.
7. Verification disciplineThe anti-hallucination layer. Claim-as-candidate, ground truth beats consensus, and why “it should work now” is not a status.
8. Agentic workflowsSubagents, parallel fan-out, worktrees — and when parallelism just produces five confidently wrong answers instead of one.
9. Shipping real projectsThe discipline gap between “the demo works” and “someone is depending on this.” Rollback, verify-after-deploy, the safety laws.
10. The capstoneBuild one real, complete piece of software end to end, using every layer from 1–9 deliberately instead of by accident.

Plus four hands-on exercises — write your first law, build a hook, build a skill, verify a claim. Each one has the real prompt to run and the check that proves it worked. Not "try this and see" — a pass condition.