Claude Code Mastery

Introduction

Course goals, audience from beginner to experienced dev.

TakeawayStay in the loop as the AI codes.

Installing & setting up

Install Claude Code, auth, trust folder, verify with the CLI.

TakeawaySetup is the hardest part; everything after is simpler.

Talking with code

Explore an open-source codebase with scoped, constrained questions.

TakeawayFrame questions explicitly and state your level.

Exercise: asking questions

Three targeted questions about a real codebase at your experience level.

TakeawayTelling Claude your level prevents wasted answers.

Models, context & usage

Tokens, 200k/1M context, /clear, Opus/Sonnet/Haiku, effort, /usage.

TakeawayReset context periodically; watch your usage ceiling.

First coding task

CSV invoices to PDFs with @filename, permissions, and verification.

TakeawayClear prompts with constraints; let Claude handle mechanics.

Vibe coding vs agentic engineering

Vibe coding for one-offs; agentic engineering for serious projects.

TakeawayFor real projects, document decisions and let agents enforce them.

CLAUDE.md

Per-project and .local files, auto-appended, nested per directory.

TakeawayCLAUDE.md travels across conversations and enforces conscious decisions.

Building your CLAUDE.md

/init to seed then trim; merge examples; add project rules.

TakeawayCLAUDE.md is personal and project-specific; treat it as code.

CLAUDE.md in practice

Migrate JS to TS with unit tests under new CLAUDE.md rules.

TakeawayCLAUDE.md guides agents to ask before acting.

Voice input

/voice, Super Whisper, Whisper for transcription.

TakeawayVoice captures thinking better than bullet points.

Claude and git

Git basics with Claude doing the heavy lifting; worktrees introduced.

TakeawayWorktrees let agents build multiple features at once.

Using git with Claude

git init, GitHub SSH, paste docs as markdown, voice-drive git.

TakeawayClaude can run setup commands; do not do it manually.

Tool permissions

Allow/ask/deny per tool; bash scoping; shared vs .local settings.

TakeawayDeny reads to .env, secrets, and production databases.

What to commit

Commit CLAUDE.md, shared settings, subagents; ignore .local.

TakeawayShare team settings; keep local overrides out of the repo.

Co-author commits

Claude as commit co-author; includeCoAuthoredBy; rebase to clean history.

TakeawayOne setting strips Claude's co-author tag everywhere.

Plan mode

Plan first, build after; review and push back before commit.

TakeawayYou control direction before code is written.

Using plan mode

Shift+Tab toggles; auto-accept or review; when planning pays off.

TakeawayPlanning costs upfront and saves time fixing mistakes.

Subagents

Isolated workers, summaries back to main context; built-in and custom.

TakeawayDelegate isolated chunks so the main agent stays focused.

Build a review subagent

Read-only reviewer scoped to uncommitted code; markdown + frontmatter.

TakeawayOne job per subagent.

What is a skill?

Reusable prompts in .claude/skills/; command or auto trigger.

TakeawaySkill = what should be done; subagent = where it runs.

Build a command skill

Build a skill with Claude; name, command, restart to load.

TakeawaySkills let you run a good prompt again on demand.

Existing skills

skills.sh marketplace; audit before install.

TakeawayOnly use trusted skills; check the security audit first.

Using skills in design

Install via CLI; auto-invocation; frontend-design skill demo.

TakeawayA good skill changes output quality dramatically.

Hooks

Non-negotiable rules on Claude Code events; block bad commits.

TakeawayPrompts are preferences; hooks are things that must be true.

Plugins

Bundle skills, hooks, subagents, MCP; install from folder or marketplace.

TakeawayA plugin is a full workflow packaged as one install.

Permission modes

Six modes from default to bypass; Shift+Tab to switch.

TakeawayStart restrictive; bypass is last resort.

Bypass / YOLO

--dangerously-skip-permissions; prompt injection risk; containers only.

TakeawayBypass mode is dangerous; use inside containers only.

Auto mode

AI classifier blocks risky bash; respects in-chat rules.

TakeawayAuto mode trades prompts for AI-based safety.

Dontask mode

Deny by default; allow only via explicit rules; for CI.

TakeawayDontask inverts permissions: allow by explicit rule.

Sandboxing

Filesystem and network wall; blocks compromised packages.

TakeawaySandboxing creates a hard wall bash cannot escape.

Git worktrees

claude -w branch checks out parallel folders; auto-cleanup.

TakeawayWorktrees let agents work on separate features in parallel.

Two agents at once

Multiple terminals, worktrees, auto mode, orchestrator pattern.

TakeawayMultiple agents can build separate features at once.

Capstone: joining a project

Technical onboarding; settings.json; CLAUDE.md vs agents.md.

TakeawayOnboarding first; do not /init over an existing project.

Setting up the environment

Testing setup first; agents.md rules; review plugins and skills.

TakeawayTest setup first; agent rules in agents.md, DRY.

Setting up hooks

Stop hook runs unit tests; file-change hook lints changed files.

TakeawayTests are mandatory; linting and e2e are advisory.

Build the feature

Ship missing feature; subagents; tests as hard gate, review as advisory.

TakeawayTests are gates; reviews are suggestions.

ADRs

docs/adr/ with index; document why for hard-to-reverse decisions.

TakeawayADRs document why, not what; agents read them before touching foundations.