32 Tricks to Level Up Claude Code in 16 Mins

This guide outlines a comprehensive set of "Cloudcode" hacks, categorized from beginner to advanced, designed to optimize workflow, improve code quality, and maximize the efficiency of AI-assisted development.

Beginner Hacks: Foundation and Workflow

• /init: Run this on every project to generate a cloud.md file. This acts as a project "cheat sheet," mapping architecture and conventions so the AI maintains context across sessions.
• Status Line: Use /status line to create a terminal dashboard displaying model, context usage, and cost, helping to prevent "context rot."
• Voice Input: Utilize native /voice commands or third-party dictation apps to code via natural language.
• Context Management: Keep context small by breaking large tasks into granular steps. Use /context to identify "token bloat" (e.g., system prompts or unnecessary file contents) and prune it.
• Compaction & Clearing: Use /compact at 60% context usage to compress history while preserving critical data (e.g., API/schema decisions). Use /clear to reset for new tasks while retaining the cloud.md file.
• Plan Mode: Always start in "Plan Mode" (Shift+Tab). This forces the AI to outline steps and ask clarifying questions before writing code, significantly reducing revision cycles.
• Junior Developer Mindset: Instead of direct commands, present problems (e.g., "How should we handle growth tracking?") to allow the AI to reason through the approach.
• Verification Loops: Build self-checking into to-do lists. Require the AI to take screenshots or run Chrome DevTools to verify functionality before marking a task as complete.

Intermediate Hacks: Scaling and Automation

• Sub-agents: Deploy isolated sub-agents for parallel tasks. This keeps the main thread clean and allows for model optimization (e.g., using the cheaper "Haiku" model for sub-tasks and "Opus" for the main thread).
• Custom Skills: Create reusable prompt files in the .cloud/skills directory (e.g., codereview.md). These can be committed to GitHub for team-wide standardization of SOPs.
• cloud.md Optimization: Keep this file lean (150–200 lines). Use it to route the AI to external documentation or style guides rather than pasting full content, which saves tokens.
• Steering: Use /rewind to undo mistakes and "exit early" by hitting escape if the AI heads down the wrong path.
• Aggressive Feedback: If an output is mediocre, demand a different approach. Once a better result is achieved, instruct the AI to update its cloud.md or skill files to prevent future errors.
• Hooks & Visuals: Use /hooks for audio notifications upon task completion. Leverage screenshots to allow the AI to perform visual quality assurance on UI layouts.
• Chrome DevTools: Use the browser tool to interact with apps, fill forms, or test functionality, effectively bypassing the need for explicit APIs in some scenarios.
• Cloning Inspiration: Feed screenshots of preferred designs to the AI to replicate patterns without generating "generic AI slop."

Advanced Hacks: Power User Strategies

• Git Worktrees: Use cloud --worktree [feature-name] to run multiple parallel sessions on different branches without overwriting files.
• API Endpoints vs. MCP: If token usage is critical, hardcode specific API endpoints instead of loading full Model Context Protocol (MCP) server definitions.
• Recurring Tasks: Use /loop for background monitoring (e.g., checking deployments every 5 minutes). Note: These last up to 3 days.
• Remote Hosting: Run Cloudcode on a VPS for "always-on" sessions, allowing interaction via SSH or Telegram.
• Remote Control: Control local sessions from a mobile device or browser, allowing for "on-the-go" development.
• Data Analytics: Connect CLI tools (e.g., BigQuery’s bq) to perform natural language data analysis without writing SQL.
• Ultra Think: Use this mode for complex architectural decisions; it allocates a maximum "thinking budget" of ~32,000 tokens before responding.
• Safe Autonomy: Instead of "dangerously skip permissions," configure an allow/deny list to automate safe tasks while blocking destructive commands (e.g., delete).
• Agent Teams: Unlike sub-agents, these agents share a task list and communicate with each other, providing a more cohesive output for large-scale projects.
• Context 7 MCP: Install this to pull real-time, version-specific documentation for libraries like Next.js or React, bypassing the AI’s training data cutoff.

Key Concepts

• Context Window: The amount of information the AI can "hold" at once; managing this is critical for performance and cost.
• Cloud.md: The central configuration file for project-specific context and rules.
• MCP (Model Context Protocol): A standard for connecting AI assistants to external data and tools.
• Sub-agents: Isolated AI instances that handle specific, parallelized tasks.
• Token Bloat: The accumulation of unnecessary data in the context window that degrades performance and increases costs.
• Plan Mode: A non-execution state where the AI researches and outlines steps before coding.
• Git Worktrees: A Git feature allowing multiple branches to be checked out simultaneously in different directories.