Skip to main content

Claude Code

Claude Code is Anthropic's terminal-based coding agent. You run it in your project directory, describe what you want in natural language, and it reads files, edits code, runs commands, and iterates — asking permission before actions that change your system.

# Install (Node 18+)
npm install -g @anthropic-ai/claude-code

# Run it inside your project
cd my-project
claude

It also has VS Code / JetBrains extensions and a web app, but the CLI is the core experience.

Using Claude optimally​

The difference between a frustrating session and a great one is mostly about context and workflow, not the model.

Give it a CLAUDE.md​

Create a CLAUDE.md at your repo root (run /init to generate one). Claude reads it automatically at the start of every session. Put in it the things you'd tell a new teammate:

  • How to build, test, and run the project.
  • Project conventions (style, directory layout, "always run the linter").
  • Gotchas and commands it should prefer.

Keep it concise — it is loaded into context every time. You can also keep a personal ~/.claude/CLAUDE.md for preferences that apply across all projects.

Plan before large changes​

For anything non-trivial, ask Claude to make a plan first (there is a plan/read-only mode). Review the plan, correct wrong assumptions, then let it implement. Course-correcting a plan is far cheaper than undoing a bad implementation.

Be specific​

Vague prompts get vague results. "Fix the bug" is weak; "the parse() function returns null for empty input — it should throw InvalidArgument; add a unit test" is strong. Point it at specific files, paste error messages, and describe the expected behavior.

Manage context​

  • Use /clear between unrelated tasks so old context doesn't pollute the new one (and to save tokens).
  • For big tasks, have it work in stages and verify each stage.
  • Reference files/symbols explicitly so it doesn't have to search blindly.

Let it verify its own work​

Claude is most reliable when it can check itself: run the test suite, run the compiler, run the program. Tell it how to run your tests and ask it to keep going until they pass. A tight edit → build → test loop catches its own mistakes.

Use subagents and parallelism​

For large or independent chunks of work, Claude can spawn subagents to explore or work in parallel. This keeps the main conversation focused and covers more ground.

Connect tools with MCP — carefully​

Claude Code is an MCP client. Adding servers (GitHub, a database, a browser) makes it far more capable, but read the MCP page first — connecting tools has real security implications.

Version control is your safety net​

Commit early and often so you can always git diff what Claude changed and roll back if needed. Consider working on a branch (or a git worktree) so experiments stay isolated.

--dangerously-skip-permissions​

By default, Claude Code asks for approval before actions that could affect your system — editing files, running shell commands, deleting things, making network requests. This human-in-the-loop check is the main guardrail.

The flag --dangerously-skip-permissions (informally "YOLO mode") disables those prompts. Claude will edit files and run commands without stopping to ask.

claude --dangerously-skip-permissions

Pros​

  • No interruptions. Long, mechanical tasks (large refactors, fixing hundreds of lint errors, scaffolding) run start-to-finish without you babysitting each step.
  • Good for automation. Useful in CI or scripted/batch runs where no human is present to click "approve".
  • Faster iteration on throwaway work where the blast radius is contained.

Cons / risks​

  • It can run anything. A wrong turn — a bad rm -rf, an overwritten file, a force-push, an unwanted apt install — happens with no chance for you to intervene.
  • Prompt-injection amplifier. If Claude reads untrusted content (a web page, a third-party issue, a malicious dependency) while permissions are off, injected instructions can execute unchecked. See the "lethal trifecta" note on the MCP page.
  • Harder to review. You lose the natural checkpoint where you'd notice "wait, why is it touching that file?"

Recommendation​

Keep permission prompts on for normal work on anything you care about. Only use --dangerously-skip-permissions when the environment is isolated and disposable — a container, a VM, or a throwaway git worktree with no access to secrets or production systems — and ideally when the task doesn't involve untrusted external input. If you find the prompts tedious, prefer narrowing permissions (allowlisting specific safe commands) over turning them off entirely.