Step 5: Advanced Features (20 minutes)

These are power-user features. You don’t need them on day one, but they’ll multiply your productivity as you get comfortable.

Permission Modes

Claude Code asks permission before editing files and running commands. You can control this:

Mode	Command	Best For
Default	`claude`	Learning, unfamiliar projects
Auto-edit	`claude --auto-edit`	Daily coding (recommended)
Plan	`claude --plan`	Read-only review sessions
Full auto	`claude --dangerously-skip-permissions`	CI/CD only — never on your machine

For granular control, add an allowlist to ~/.claude/settings.json — allow edits and safe commands, but still require approval for git push, rm, and destructive operations.

See the full permissions guide for allowlist patterns and recommendations by role.

Model Routing

Claude Code has three model tiers. The default (Sonnet) handles most work, but you can optimize:

Model	When to Use	How
Haiku	Quick lookups, formatting	`/fast` toggle
Sonnet	Standard coding (default)	Automatic
Opus	Complex architecture, hard bugs	Use when Sonnet struggles

Rule of thumb: If Sonnet can’t solve it in 2 attempts, try with Opus reasoning.

See the full model comparison guide.

Context Management

Your context window is finite. Manage it:

Context %	Action
< 50%	Keep working normally
50-80%	Run `/compact` to compress
> 80%	Start a fresh session (use `/handoff` first)

Check usage anytime with /context.

Why this matters: Claude gets worse as context fills up. Old failed approaches, debug output, and rejected ideas compete for attention. Fresh sessions produce better code.

Multi-Agent Work

For large features (5+ files), Claude can spawn parallel agents:

Build a notifications system:
- Database schema and migration
- API endpoints (GET list, PATCH mark-read)
- React component with dropdown

These are independent — parallelize where possible. Cap at 3 agents.

Safety rules:

Cap at 3-4 agents (more = rate limit issues)
Each agent works in its own worktree (no conflicts)
Always run the full test suite after merging agent results
Specify model: "sonnet" for implementation agents

See the multi-agent example for a full walkthrough.

Hooks In-Depth

Hooks are automatic checks. You’ve been using them — here’s what’s happening:

Hook	When It Runs	What It Catches
`ts-check.sh`	After every TS file edit	Type errors
`lint-check.sh`	After every JS/TS edit	ESLint violations
`format-check.sh`	After every edit	Formatting issues
`env-guard.sh`	Before git commands	Secrets in commits
`pre-commit-guard.sh`	Before git commit	console.log, debugger
`commit-message-check.sh`	Before git commit	Non-conventional commit messages
`test-on-save.sh`	After edits (async)	Related test failures
`build-check.sh`	After TS edits	Build failures, OOM
`session-start-check.sh`	Session start	Environment issues

You can disable individual hooks by removing them from ~/.claude/settings.json.

Reverse Prompting

Instead of writing detailed specs, let Claude ask you:

I need to add user authentication to this app.
Ask me 20 clarifying questions before you start.

Claude’s questions reveal edge cases you haven’t considered. Answer them, and you get a better spec than you’d write yourself.

Custom Skills

You can create skills specific to your team’s workflows:

Create ~/.claude/skills/my-skill/SKILL.md
Add YAML frontmatter with name, description, metadata
Write the step-by-step workflow in markdown
Use it with /my-skill

See CONTRIBUTING.md for the format.

Topic	Resource
All 27 skills	Workflow decision tree
Common mistakes	22 anti-patterns
Model selection	Model comparison
MCP servers	MCP guide
Prompt patterns	22 prompt patterns
Full reference	Power user guide (667 lines)
Agent Teams	Agent Teams guide
CI/CD automation	GitHub Actions guide
Enterprise setup	Enterprise Governance