Claude Code for Builders: the advanced Claude Code course

You ship: A setup you can trust unattended: hooks that stop risky actions before they happen, so you stop babysitting every command.

1. 1.1Terminal vs desktop, and a tour of my setup. When to use the terminal and when the desktop app, then a tour of James's own setup: output styles, settings scopes, the HUD plugin and a custom status line.
2. 1.2Power commands. The commands that change how a session runs, from /context, /usage and /doctor to effort levels, /compact and /resume.
3. 1.3Hooks: rules Claude can't talk its way out of. Claude Code hooks that block risky actions before they happen, what makes a good one, and how to test your hooks against bypasses.
4. 1.4Permissions and auto mode. Ask and deny rules, auto mode, plan mode and sandboxing, so Claude can work unattended inside limits you set.
5. 1.5Plugins and a versioned setup. Package your setup as plugins and keep ~/.claude in git, so it travels with you between machines.

You ship: A real feature shipped end to end with James's exact method, so your agents build the right thing first time.

1. 2.1The loop, and research before deciding. The compound engineering loop, and researching with parallel agents, Perplexity and Firecrawl before you decide anything.
2. 2.2The PRD method (with the interview inside it). Turn a rough idea into a written PRD by letting Claude interview you, with "done" defined as a test.
3. 2.3Adversarial review of the plan. Have agents attack the plan before a line of code is written, while changes are still cheap.
4. 2.4See it, mark it up, build it. Three wireframes side by side, a DESIGN.md file and annotated screenshots, so the agent builds what you pictured.
5. 2.5Build, test, ship. The worked example built end to end, with automated tests, one PR per milestone and Claude checking its own UI in a browser before it says done.

You ship: Several agents building one feature at once, while you get on with something else.

1. 3.1Subagent, skill or custom agent?. When each one fits, and why you mostly don't need custom agents: a skill run in a subagent covers most jobs.
2. 3.2Many agents at once. Background agents, agent teams and a pinned model per subagent, so several Claude Code subagents run in parallel.
3. 3.3/batch and workflows. Fan one job out across many agents with /batch and workflows.
4. 3.4Research and critic loops. Parallel research subagents and critic loops that pressure-test a piece of work until it holds up.

You ship: Skills that measurably improve themselves, so your best workflows get better every time they run.

1. 4.1Skills done properly. Skill frontmatter, /skill-doctor, and when a plugin is the better package than a skill.
2. 4.2Skill evals. Evals, LLM judges and held-back tests that tell you whether a skill actually got better.
3. 4.3Hill-climbing: quality, then cost. Improve a skill against its evals, then find the cheapest model that still passes them.
4. 4.4Scanning skills before you trust them. Check third-party skills before you install them, and add a hook that scans every new one.

You ship: A second brain every session starts from, so Claude already knows your business, your rules and your past decisions.

1. 5.1Docs over memory. Why context belongs in files you own and can read, not in hidden memory.
2. 5.2CLAUDE.md best practices. A lean CLAUDE.md, a context budget audit, AGENTS.md for every tool and CLAUDE.md imports.
3. 5.3Sources of truth and path-scoped rules. A source-of-truth table, and rules that load only when Claude touches the files they cover.
4. 5.4One context for the whole team. Share one context across a team, so everyone's agents start from the same knowledge.
5. 5.5Connect everything to Claude Code. MCP, CLI or API: how to connect each of your tools, so Claude can read and act on your whole business.

You ship: PRs, agent review and preview deploys, so nothing reaches production without your OK.

1. 6.1GitHub 101. The GitHub basics agents rely on, and how Claude drives PRs, checks and merges with the gh CLI.
2. 6.2Git worktrees. Git worktrees so parallel agents never collide: one task, one branch, one PR.
3. 6.3Agentic code review. Adversarial and cross-model code review, plus a security pass on the code your agents write.
4. 6.4Nothing ships without your OK. Push-to-main blocks, PR holds and secret guards, so production is the one place agents stop and ask.
5. 6.5Preview, deploy, watch. Preview deploys, Vercel from the terminal, and error reporting so you hear about problems first.

You ship: Claude Code running 24/7 on a machine that never sleeps, steered from your phone.

1. 7.1A machine that never sleeps. A Mac mini or a VPS running Claude Code around the clock, reached securely over Tailscale and tmux.
2. 7.2Drive it from anywhere. Remote Control, cloud sessions and your phone, so the work carries on when you leave your desk.
3. 7.3Routines. Claude Code routines on a schedule, from /loop to API and GitHub triggers, with rules that keep the cost down.
4. 7.4Routines that improve themselves. Routines that read their own run logs and get better each time they run.
5. 7.5A setup that audits itself. Subagents that audit your whole setup every week and tell you what to fix.

You ship: Routine work routed to cheaper models inside Claude Code, with your subscription kept for what matters.

1. 8.1OpenRouter, subscription intact. Run OpenRouter models inside Claude Code without breaking your Claude subscription. Read the free guide to OpenRouter in Claude Code.
2. 8.2Kimi, GLM and other open models. Where open models such as Kimi and GLM do the job well enough, and how to switch to them.
3. 8.3Ollama in the desktop app. Run local models through Ollama in the Claude Code desktop app.
4. 8.4Model routing for subagents and background calls. Send subagents and background calls to cheaper models, and keep the best model for the work that needs it.