Agent Cookbook

Beginner

I've Been Using Claude Code Wrong for Months—These 40 Tips Completely Transformed My Workflow

Starting from common pitfalls, systematically organize 40 tips for using Claude Code efficiently, covering configuration, context management, automation, CLAUDE.md conventions, multi-agent concurrency, and other core scenarios.

Most people use Claude Code as an autocomplete tool. The top 1% use it like an operating system to command the entire development workflow.
"The gap between average users and experts isn't technical skill—it's configuration. That gap is worth 4–6 hours per week."
Here are 40 Claude Code best practices that truly transform your workflow.

1. Basic Configuration#

01 — Set a cc Alias to Skip Confirmation Every Time#

Add this to your ~/.zshrc or ~/.bashrc:
bash
alias cc='claude --dangerously-skip-permissions'
Then run source ~/.zshrc. After that, typing cc will skip the permission confirmation prompt for every operation. This assumes you understand what you're authorizing—it's fast because it trusts you.

02 — Enable the Status Bar for a Real-Time HUD#

Run this inside Claude Code:
/statusline
It generates a shell script that displays, in real-time at the bottom of your terminal after each AI response: the current directory, branch name, and context usage. It's like a session-level HUD.

03 — Extend the Context Window to 1 Million Tokens#

Sonnet 4.6 and Opus 4.6 support a 1 million token context. Switch mid-session with:
/model opus[1m]
Start with 500k and gradually increase to find the optimal balance before compression.

04 — Configure Output Style Once, Apply Forever#

Run /config and choose a style: Explanatory, Concise, or Technical. You can also create fully custom style files in ~/.claude/output-styles/. Once configured, you won't need to manually correct the AI's tone every time.

05 — Control Claude Code Remotely from Your Phone#

bash
claude remote-control
Then connect from claude.ai or the mobile app. Initiate a large refactor on your PC and check progress from your phone anytime—it feels like science fiction.

2. Workflow Speed Tips#

06 — ! Prefix: Pipe Bash Output Directly into Context#

Type !git status or !npm test, and the output lands directly in Claude's context without copy-pasting, dramatically improving the debugging loop.

07 — Esc to Pause, Esc+Esc to Rewind#

  • Esc: Immediately abort Claude's current operation
  • Esc+Esc (or /rewind): Opens a menu to restore code, conversation, or both
This is your "undo" button. Use it boldly for ideas you're only 40% sure about.

08 — Ctrl+S: Save Your Current Draft#

Need to ask a quick question mid-draft? Ctrl+S saves the draft; after asking, the draft auto-restores. A lifesaver for complex multi-step tasks.

09 — Ctrl+B: Run Time-Consuming Tasks in the Background#

While Claude runs tests or builds, press Ctrl+B. It continues working in the background while you plan the next steps. True parallel productivity.

10 — /btw: Ask Side Questions Without Polluting the Main Conversation#

/btw opens a floating layer for quick questions like "Why use this method?" The main conversation history stays clean, but your curiosity is satisfied.

11 — Ctrl+G: Edit Claude's Plan Before It Acts#

After Claude presents a plan, Ctrl+G opens the plan in an editor. Adjust steps, correct direction, then let it execute. Avoid discovering 20 minutes later that you're off track.

12 — Voice Input Produces Better Prompts Than Typing#

/voice
Enable push-to-talk mode. When speaking prompts, you naturally add more context, constraints, and details, resulting in significantly better output than brief typed commands.

3. Context and Prompt Management#

13 — Execute /clear Between Different Tasks#

A clean session is always better than a chaotic 3-hour session. Accumulated context silently dilutes your instructions. /clear is key to preventing output quality degradation throughout the day.

14 — If You've Corrected Twice and It's Still Wrong, Start a New Session#

If you've corrected Claude twice and it's still off, don't try a third time—/clear and rewrite the prompt, incorporating the lessons learned. Continuing a broken conversation only adds confusion.

15 — Don't Describe Bugs; Paste Raw Data#

Instead of explaining a bug, paste the raw error log, CI output, or Slack screenshot, then say "Fix it":
bash
cat error.log | claude "Explain this error and provide a fix"
Abstract descriptions lose details; raw data yields precise results.

16 — Use Plan Mode for Architectural Changes#

Press Shift+Tab to enter Plan Mode. Essential for multi-file changes or structural adjustments. The upfront planning overhead prevents Claude from spending 20 minutes confidently solving the wrong problem.

17 — Directly Specify Files to Save Tokens#

Use @src/auth/middleware.ts to reference a file directly. Claude parses it automatically, avoiding a full codebase search for context you already know.

18 — Explore Unfamiliar Code with Vague Questions#

"How would you improve this file?" can uncover inconsistencies you didn't know to ask about. When exploring unfamiliar territory, vagueness is a feature.

19 — Provide Retention Guidance When Compressing#

When using /compact, tell Claude what to keep: "Focus on retaining API changes." Compression without guidance loses the context thread; guided compression distills precisely.

20 — ultrathink Unlocks Adaptive Reasoning#

Append ultrathink to any prompt (Opus 4.6 only):
Refactor this payment module, ultrathink
Claude dynamically allocates reasoning budget based on the actual complexity of the problem. For genuinely difficult problems, the quality improvement is noticeable.

4. Automation, Tools, and MCP#

21 — Have Claude Self-Check Results#

Include a test command in your prompt:
Refactor the auth module, run the test suite, fix any failing cases, then tell me it's done
This single instruction can yield 2–3x quality improvement because it creates a feedback loop.

22 — Install the LSP Plugin for Automatic Diagnostics#

/plugin install typescript-lsp@claude-plugins-official
The LSP plugin lets Claude automatically fetch diagnostics after every edit, fixing type errors before you even notice them.

23 — Prefer CLI Tools Over MCP Servers#

CLI tools use less context than MCP Servers. Teach Claude to use gh for PRs and sentry-cli --help for debugging production issues. The saved context accumulates over long sessions.

24 — Prioritize Installing These Four MCP Servers#

The four with the best signal-to-noise ratio:
  • Playwright — UI verification
  • PostgreSQL/MySQL — Schema queries
  • Slack — Directly read bug discussion threads
  • Figma — Convert designs to code
Master these four before expanding further.

25 — /loop for Periodic Background Checks#

/loop 5m check if deploy succeeded
Set up a background loop check. Keep the session open, continue working, and Claude proactively reports when there's news.

26 — Whitelist Trusted Commands with /permissions#

Stop manually approving npm run lint every time. Whitelist trusted commands to eliminate the hidden efficiency loss from frequent confirmations.

5. Mastering CLAUDE.md and the Rules System#

27 — Generate with /init, Then Cut It in Half#

Run /init to generate an initial CLAUDE.md, then ruthlessly delete anything you can't justify keeping. Every extra line of instruction consumes tokens and quietly reduces attention quality elsewhere.

28 — The Golden Test for CLAUDE.md#

For every line, ask yourself: "Without this, would Claude make a mistake?" If the answer is no → delete it. You have a budget of roughly 150–200 instructions before compliance starts to decline. Use it wisely.

29 — Auto-Update Rules After Every Mistake#

When Claude makes an error, say:
Update CLAUDE.md to ensure this problem never happens again
Your rules file becomes a living system that evolves with every session.

30 — Use .claude/rules/ for Conditional Rules#

Add a path declaration at the top of a rules file:
yaml
---
paths:
  - "src/api/**"
---
TypeScript rules only apply to .ts files; database rules only apply to the /db directory. Context is precise, no waste.

31 — Keep CLAUDE.md Lightweight with @imports#

Reference @docs/git-instructions.md instead of pasting the content directly. Claude reads it on demand, keeping the base context lightweight.

32 — Skills: On-Demand Extended Knowledge#

Skill files in .claude/skills/ are loaded only when invoked, consuming no context otherwise. Think of them as library files—mounted when needed, invisible when not.

33 — CLAUDE.md for Suggestions, Hooks for Mandatory Rules#

Claude follows CLAUDE.md about 80% of the time. For non-negotiable requirements like formatting, security, or code conventions, use Hooks—they execute every time, no exceptions.

34 — PostToolUse Hook: Auto-Format After Every Edit#

Add this to .claude/settings.json:
json
"hooks": {
  "PostToolUse": [{
    "matcher": "Edit|Write",
    "hooks": [{
      "type": "command",
      "command": "npx prettier --write \"$CLAUDE_FILE_PATH\" 2>/dev/null || true"
    }]
  }]
}
Automatically runs Prettier after every edit, eliminating manual formatting entirely.

35 — PreToolUse Hook: Intercept Dangerous Commands#

Intercept Bash tool execution to block dangerous operations like rm -rf or DROP TABLE. The PreToolUse Hook is the safety net that lets you give Claude more autonomy.

6. Advanced: Worktree, Agents, and Isolated Execution#

36 — --worktree for Parallel Branch Development#

bash
claude --worktree feature-auth
Creates an independent working copy. Run 3 sessions simultaneously for different features without interference, directly multiplying throughput.

37 — Use Sub-Agents to Keep the Main Context Window Clean#

Use a sub-agent to figure out the payment flow
This spawns an independent instance to read relevant files and return a summary—your main context stays lean and focused.

38 — Create Custom Sub-Agents for Frequent Tasks#

/agents
Save pre-configured agents in .claude/agents/: a fast search agent based on Haiku, a strict TypeScript review agent, a documentation writing agent. Your dedicated AI team, on call.

39 — Agent Teams: Large-Scale Parallel Work#

Enable the experimental feature:
bash
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 claude
One lead agent distributes tasks to 3–5 parallel sub-agents. For large-scale research or multi-module refactoring, this is speed on another level.

40 — /sandbox: A Safe Sandbox for Unsupervised Experiments#

/sandbox
Achieves OS-level isolation via Seatbelt or bubblewrap, letting Claude run wild on experimental refactoring with zero risk to your real system. Review the diff afterward and merge what you like.

Summary#

"Most engineers optimize code. The fastest ones optimize their AI workflow. These are two completely different compounding advantages."
These 40 tips require no new subscriptions, and you don't need to be a 10x engineer. They're all hidden in the tools you already use, waiting to be discovered.
Pick 5 tips from this list and implement them this week. Come back in 30 days—your workflow will be fundamentally transformed.
Previous
Understanding Hermes Supervisor through the Evolution of Anthropic's Harness
Next
Harness design for long-running application development