Claude Code Cheatsheet: Commands, Hooks & Subagents

Most developers use Claude Code like a fancy autocomplete. They type a request, read the answer, and move on. That's maybe 20% of what the tool can do.

Claude Code is a terminal agent with full project access. It reads files, runs shell commands, commits to git, delegates work to parallel subagents, and enforces rules through lifecycle hooks. Once you know the full feature set, the way you work changes. This cheatsheet covers everything: every slash command, every shortcut, CLAUDE.md, hooks, subagents, and PR review — with working examples throughout.

Claude Code is Anthropic's official CLI AI assistant. Unlike IDE-embedded tools, it runs in your terminal with full access to your project: it reads and writes files, executes shell commands, interacts with git, and connects to external services through MCP servers.

Installing it takes one command.

/init reads your project structure and generates a CLAUDE.md file plus a .claude/ directory. That CLAUDE.md becomes Claude's instruction manual for every future session in that project. Get that file right and the rest of your sessions will be significantly more consistent.

Type / on an empty prompt and press Tab to see all available commands, including any custom ones you have defined. Here is the full built-in set.

Four of these commands deserve a bit more explanation because they change how you run longer sessions.

The ! prefix is underused. Typing !npm test pulls the test output directly into context so Claude can reason about failures without you copy-pasting. Ctrl+V for screenshots works especially well for sharing browser error states or design references.

CLAUDE.md is a Markdown file at the root of your project (or in .claude/) that Claude reads at the start of every session. It's your project's standing instructions: stack details, naming conventions, commands to run, rules to follow.

Here is a practical example for an API project.

Keep CLAUDE.md under 500 lines. When rule sets grow large, split them into separate files under .claude/rules/ and reference them from the main CLAUDE.md — Claude follows imported files too.

The Commands section matters most. Claude references it every time it needs to run a linter or test suite. If your project has non-standard script names, see our guide to npm scripts — consistent naming makes Claude significantly more reliable without extra instruction.

CLAUDE.md instructions are advisory. In practice, Claude follows them roughly 80% of the time. For rules that must never be broken — blocking dangerous shell commands, enforcing formatting — use hooks instead. CLAUDE.md is for preferences and context; hooks are for enforcement.

Hooks are shell scripts that run at fixed points in Claude's execution lifecycle. Unlike CLAUDE.md instructions, hooks are deterministic. If a hook exits with a non-zero status code, Claude stops and surfaces the error. There is no workaround.

There are five hook types.

Hooks live in .claude/settings.json. Here is a PostToolUse hook that runs Prettier automatically every time Claude writes a file.

The matcher field targets a specific tool — Write in this case. $CLAUDE_TOOL_INPUT_FILE_PATH is an environment variable Claude populates with the file path being written.

For safety-critical enforcement, PreToolUse is the right hook. This example blocks any rm -rf command before it executes.

When the hook exits with code 1, Claude halts and reports the failure. The file system stays untouched.

PostToolUse hooks complement Git commit hooks rather than replacing them. If your project already uses Husky and lint-staged for pre-commit enforcement, Claude Code hooks catch format issues at the AI write layer before code ever reaches a commit.

Claude Code lets you extend the slash command system in two ways. Custom commands are simple Markdown prompt files. Skills are structured, multi-file workflows that Claude can invoke automatically when the situation calls for them.

A custom command for a standup summary is just a Markdown file with a prompt. It shows up as /standup in the command list and runs that prompt verbatim when invoked. Create it by adding .claude/commands/standup.md with your prompt text inside.

A skill makes sense when your workflow has branching logic, depends on supporting templates, or is something Claude should recognize and apply on its own without you typing a command. The complexity threshold is real: if you find yourself maintaining a multi-paragraph custom command with conditional instructions, it belongs in a skill.

Context windows fill up. In a long session, the earlier parts of your conversation fade from Claude's working memory, and output quality drops. Subagents solve this by offloading work to separate Claude instances with clean, empty context.

Your main session delegates a task to a subagent. The subagent works in isolation, completes the task, and returns a summary. The main session only sees the summary — not the thousands of lines of intermediate output the subagent generated.

The bigger win is parallelism. Claude can run three subagents simultaneously. You can review the auth module, run API integration tests, and check a DB migration at the same time rather than sequentially. On tasks that would take 20 minutes in a single session, parallel subagents regularly cut that to 7.

Custom subagents let you pre-configure a Claude instance for a specific role. Define one in .claude/agents/name.md with its own model choice, allowed tools, and standing instructions. A dedicated test-runner subagent configured to only use lightweight models and only run read commands is cheaper and safer than an unconfigured general-purpose instance.

Claude Code integrates with GitHub for pull request review. Set it up once and you can run reviews directly from the terminal with findings posted as inline GitHub comments.

For CI integration, /autofix-pr triggers an automated fix attempt when CI fails. Claude reads the failure logs, patches the issue, and pushes a follow-up commit to the same branch.