Capstone Projects — Prove Each Level

Reading ≠ skill. Each capstone below exercises everything in its stage. Do them in a real repo (your own project, or clone any open-source app you like). Don’t move to the next stage until the current capstone works.

✅ Capstone 1 — Foundations (after files 00–06)

“Tour guide”: In a repo you’ve never read:

Have Claude produce an architecture summary in ARCHITECTURE.md
Find and fix one real bug or code smell it identifies
Commit with a Claude-written message on a new branch
Exit, then resume the same session with claude -c and ask a follow-up

You’ve proven: install, prompting, permissions basics, git flow, resume.

✅ Capstone 2 — Working Sessions (after files 07–11)

“Controlled demolition”:

In plan mode, plan a medium refactor (rename a core concept across files)
Approve; let Claude execute
Deliberately decide the approach is wrong → /rewind → restore code
/branch and try approach #2; keep the better branch
Run /context, then /compact with focus instructions, and finish

You’ve proven: plan mode, checkpointing, forking, context management.

✅ Capstone 3 — Configuration (after files 12–16)

“Zero-prompt test loop”: Configure a project so claude "run tests and fix failures until green" needs zero permission prompts and is still safe:

Project .claude/settings.json with permission allow-rules for your test/build commands
Sandbox enabled (WSL2 on Windows) with npm registry allowed, ~/.ssh denied
Model set per project; verify with /model and /sandbox → Config

You’ve proven: settings scopes, permission rules, sandboxing, model config.

✅ Capstone 4 — Memory & Extensions (after files 17–27)

“Team onboarding kit” for one of your repos:

CLAUDE.md < 100 lines: build/test commands, conventions, architecture pointers
Two skills: /new-endpoint (scaffolds route + test + docs) and /release-notes (generates from git log)
One hook: auto-format after every file edit (PostToolUse)
One MCP server connected (e.g. GitHub or Postgres) and used in a real task
Test: in a fresh session, /new-endpoint health should work first try

You’ve proven: memory design, skills, hooks, MCP — the whole extension layer.

✅ Capstone 5 — Agents & Parallelism (after files 28–32)

“Review fleet”:

Define two subagents: security-reviewer (read-only tools) and test-runner
Use both in one task and compare what returns to main context
Run two worktree sessions in parallel on different features
(Optional, experimental) Enable agent teams; run the three-lens PR review from 32-agent-teams.md

You’ve proven: delegation, context isolation, parallel workflows.

✅ Capstone 6 — Automation (after files 33–39)

“Hands-off pipeline”:

Headless: a shell script that pipes git diff main into claude -p and outputs JSON review findings
A routine or scheduled task: nightly dependency-audit that opens an issue when something’s outdated
GitHub Actions: @claude responds to issues in one of your repos
/code-review high as your pre-push habit for a week

You’ve proven: headless mode, scheduling, CI integration, review tooling.

✅ Capstone 7 — Pro (after files 40–43)

“Ship a real agent”: Build a small purpose-built agent with the Agent SDK (Python or TS):

Custom system prompt + one custom tool (e.g. queries your own API)
Structured JSON output; track cost per run
Deploy it somewhere real (cron on a server, GitHub Action, or a Slack-triggered flow)
Write a one-page runbook: what it does, cost ceiling, failure modes

You’ve proven: you can go from “user of Claude Code” to “builder of agents” — the actual pro bar.

Scorecard

Capstone	Done	Date	Notes
1 Foundations	☐
2 Sessions	☐
3 Configuration	☐
4 Extensions	☐
5 Agents	☐
6 Automation	☐
7 Pro	☐

🧠 Quick Recall

Trick: “A stage isn’t done until its capstone runs.”

👨‍🏫 Teach It

Do: use the scorecard as your class tracker — public progress motivates.
They’ll trip on: skipping capstones to read ahead — knowledge without reps evaporates.

Learning path: ⬅ 44-anthropic-courses.md · Index · ➡ 46-jetbrains-and-terminal-setup.md

Written by Fenil Patel