32 Claude Code Hacks in 16 Mins
A tutorial covering 32 Claude Code productivity hacks organized from beginner to advanced, including context management strategies, parallel agent workflows, and power-user features. The video progresses from basic setup commands like /init and /compact to advanced techniques like Git worktrees, agent teams, and the Context7 MCP server.
Summary
The video presents 32 Claude Code hacks organized into three tiers: beginner, intermediate, and advanced. The beginner section covers foundational workflow improvements such as running /init to auto-generate a CLAUDE.md project cheat sheet, setting up a status line dashboard in the terminal, using voice input, keeping context windows small, using /context to diagnose token bloat, and compacting or clearing sessions strategically. It also emphasizes starting every task in plan mode to reduce revision cycles, treating Claude like a junior developer by posing problems rather than issuing direct commands, prompting Claude to ask clarifying questions until 95% confident, and embedding self-verification steps directly into Claude's to-do lists.
The intermediate section introduces more sophisticated productivity techniques. These include deploying sub-agents for parallel work so the main thread stays clean, creating reusable custom skill files in the .claude/skills directory, using the cheaper Haiku model for sub-agents handling simple or high-volume tasks, continuously refreshing the CLAUDE.md file while keeping it under 150-200 lines, routing CLAUDE.md to external reference files to save tokens, exiting early when Claude goes off track, aggressively challenging mediocre outputs to get better results, using /rewind for quick rollbacks, setting up notification hooks to manage multiple concurrent sessions, using screenshots for visual self-check loops during web development, leveraging Chrome DevTools for functional app testing, and cloning inspiration sites by feeding screenshots or HTML styling directly to Claude.
The advanced section pushes Claude Code to its limits. Git worktrees allow multiple parallel isolated sessions on the same project without file conflicts. Using direct API endpoints instead of MCP servers can save significant tokens when only one specific function is needed. The /loop command enables recurring background tasks with natural language scheduling. Hosting Claude Code on a VPS enables always-on sessions accessible via SSH or Telegram. Remote control from a phone allows steering local sessions without the code leaving the local machine. Connecting CLI tools like BigQuery enables natural language data analytics without writing SQL. The 'ultra think' keyword allocates a maximum 32,000-token thinking budget for hard architectural problems. Editing permissions explicitly rather than using dangerously-skip-permissions provides safe autonomy. Agent teams, unlike sub-agents, allow all agents to communicate with each other and share task lists for more cohesive complex project outputs. Finally, the Context7 MCP server solves Claude's training data cutoff problem by injecting current, version-specific documentation from thousands of popular libraries before code generation begins.
Key Insights
- The speaker argues that using /compact with specific retention instructions — such as 'keep all API integration decisions and database schema' — allows Claude to intelligently compress conversation history while preserving only the most critical context, rather than doing a blanket compression.
- The speaker claims that having Claude run three iterative passes of build-screenshot-analyze before delivering a V1 website produces dramatically better initial results than a single-pass generation, effectively baking quality control into the build loop.
- The speaker explains that using Haiku for sub-agents doing high-volume token tasks — such as scraping hundreds of articles — while keeping the main Opus thread for synthesis and decision-making can dramatically reduce costs without sacrificing output quality where it matters.
- The speaker argues that explicitly configuring an allow list of safe commands and a deny list of destructive operations (like deletes) in Claude's permissions achieves the same speed as 'dangerously skip permissions' without the actual danger, since deny rules always take priority over allow rules.
- The speaker presents Context7 MCP as a solution to Claude's training data cutoff problem, stating it injects version-specific, live documentation from thousands of popular libraries like Next.js, React, and MongoDB directly into the conversation before Claude writes any code.
Topics
Full transcript available for MurmurCast members
Sign Up to Access