Capstone: Assemble the Full Claude Code Stack on a Real Repo
Take newsroom, a small working digest generator, and assemble every Claude Code building block from this course on it: CLAUDE.md steering, a reusable skill, a PreToolUse hook that provably blocks writes to the planted secrets file, a read-only review subagent, a headless nightly driver, and an operations runbook. This is a composition grade: the pieces must reference each other and hold up as one system. A local self-check runs your real hook and your real tests, then you submit for rubric-based feedback.
2 hrs
Est. time
4
Outcomes
6
Rubric criteria
70%
Pass score
What you'll learn
Skills you'll have real reps in after shipping this.
The scenario
You run a tiny newsroom: five story files, a digest generator, a green test suite. The digest should rebuild itself every night without you in the loop, and you want Claude Code to drive that run. Unattended means nobody is watching the session, so the configuration has to carry everything a human operator would otherwise provide: what the repo is and how it works (steering), the workflow to execute (a skill), a guard that physically cannot be talked out of protecting the credentials file (a hook), a reviewer that can read but never change anything (a subagent), a script that chains the agent run with deterministic gates (the driver), and a document the on-call person debugs from at 2 a.m. (the runbook).
This capstone is that assembly. Every component is one you built in an earlier level; what is new is the composition. The skill your driver invokes must actually exist, the runbook must name the hook file and subagent you actually shipped, and the hook must provably fire: the kit's check.py runs your real hook against simulated tool-call events and your real test suite, and writes your proof-of-work evidence only when those executable gates genuinely pass.
Your role
You are the engineer putting a small production repo on autopilot. Your deliverable is its complete Claude Code operations stack (steering, skill, hook, subagent, driver, runbook) shipped as one coherent submission, with machine-generated evidence that the guardrail fires and the tests pass.
Start the task to unlock the full brief
You'll get the step-by-step requirements, setup commands, the 6-criterion grading rubric, tips, and the ability to submit your solution for instant AI grading.
Free to start · submit when you're ready
Learning resources
Assemble the full Claude Code stack on a real repo, hands-on
This capstone has you wire every Claude Code mechanism onto one small working repo and prove the result behaves as a system. You write CLAUDE.md steering a real digest generator, package its refresh workflow as a skill, enforce a no-writes rule on the credentials file with a PreToolUse hook that a local self-check actually executes, scope a read-only review subagent, chain a headless claude -p run with deterministic test and build gates in a nightly driver script, and document the whole thing in an operations runbook. The self-check verifies the pieces reference each other: the skill your driver invokes must exist, and the runbook must name the hook and subagent you actually shipped. You come away having practiced the full configuration surface of Claude Code (steering, skills, hooks, subagents, headless automation) and the composition discipline that makes an unattended agent pipeline trustworthy.