Claude Code guide in three levels, Beginner to Advanced: hands-on lessons with checked exercises, every volatile fact cited to its source. Independent community guide.
Advanced
Advanced · about 4 hours, plus run time · Verified against Claude Code v2.1.292 (stable) on 2026-10-15
Beyond one session: parallel sessions, orchestration, unattended runs in scripts and CI, a setup your team shares, and limits that keep an autonomous run safe.
You're ready for this level if you can
- Write a project skill that runs by name and loads on its own when it applies.
- Write a hook that enforces a rule, and show it firing.
- Delegate to a custom subagent with limited tools.
- Add an MCP server at project scope, and install a plugin from an Anthropic marketplace.
- Put permission rules in the right scope, turn on the sandbox, and pick the model and effort for a task.
- Organize project memory, and find out what loaded when an instruction isn't followed.
Practice
Practice in your own repository, and run each check from your copy of the practice template as npm run check -- <lesson-id> --dir <your repo>. The GitHub Actions lesson and the capstone are the exceptions: they run in a separate template copy, which you tear down afterwards so no workflow or secret is left behind.
To go step by step, run npm run learn in your copy and use Claude Code in a second terminal beside it. The app takes you through every lesson and the capstone, one step at a time, and checks your work as you go. Each page below teaches the same lesson on its own.
A copy you made for an earlier level may not have the Advanced checks or lessons: npm run check then lists no a- lessons. Bring them in from the template, in your copy:
git fetch https://github.com/wesammustafa/claude-code-practice main
git checkout FETCH_HEAD -- checks capstone learn
git commit -m "Add the Advanced checks and lessons"
A copy older than the app has no learn script, so npm run learn says Missing script: "learn": run node learn/app.mjs instead.
Lessons
- Parallel sessions with worktrees
- Choose an orchestration pattern
- Run and save a dynamic workflow
- Script Claude Code with
claude -p - GitHub Actions
- Share your setup with a team
- Put bounds on autonomous runs
Each lesson takes about 20 minutes, and the capstone 60 to 90, plus the time an agent run takes.
Capstone
Add CI review your team can share: in a new template copy, one review skill that a bounded claude -p script and a pull request workflow with limits both run, built in two parallel worktree sessions beside a team marketplace, plus a saved dynamic workflow for a sweep before a release. Then tear down the workflow, the secret and, with an API key, the key.
By the end of this level you can
- Run parallel sessions in worktrees and merge their work.
- Choose between subagents,
/batch, a workflow or a team by need and cost, keep a session working until a check passes with/goal, and save a dynamic workflow. - Write a
claude -pscript and a GitHub Actions workflow. - Share a setup as committed config and a team marketplace plugin.
- Bound an autonomous run with least privilege, the sandbox, turn or spend limits, a review gate and a cost estimate.
Electives
Optional lessons beside the core path:
- Agent teams: a lead and teammates that message each other, staffed from your own subagent definitions.
- Agent view: hand tasks to background sessions from one screen, and step in only when one needs you.
/loopand cloud scheduled tasks: keep Claude working on an interval in your session, or on a schedule in the cloud.- Testing and publishing a plugin: check a plugin's files and behavior, then publish it in your own marketplace.
- Dev containers: run Claude Code in a container with a firewall, keep your sign-in, and let it work unattended.
Official companions
- Run agents in parallel: Anthropic's comparison of the ways Claude Code takes on several tasks at once.
- The AI-native SDLC playbook: Claude Academy's course on reshaping a team's development lifecycle around AI.
Beyond this guide
Beyond this guide: building your own agents and MCP servers, and running and rolling out Claude Code for an organization.
Sources: Run agents in parallel · The AI-native SDLC playbook