How I use Claude Code (Meta Staff Engineer Tips)

Overview

A Meta staff engineer shares 50 tips for using Claude Code effectively, emphasizing context management, keyboard shortcuts, and the strategic use of skills, MCPs, and sub-agents. He advocates for leveraging Claude Code's composability framework to build automated workflows and parallel development processes, ultimately optimizing code generation and debugging.

Key takeaways

Running Claude Code in the root directory and using `/init` ensures proper context and the creation of a `claw.md` file.
A well-structured `claw.md` file (around 300 lines) with technical architecture and validation flows improves AI coding.
Essential keyboard shortcuts like `Shift + Tab` and `Escape` enhance efficiency when using Claude Code daily.
Slash commands like `/context` and `/models` are crucial for managing context and selecting appropriate models (e.g., Opus 4.5).
Skills, MCPs, and sub-agents are composable elements that can be combined to create automated workflows.
Parallel development with multiple Claude Code instances and Git work trees enables simultaneous work on different tasks.

Chapters

0:00 Introduction to 50 Claude Code Tips

The speaker, a Meta staff engineer, uses Claude Code daily and reviews code.
The tips cover setup, cloud.md rules, and advanced techniques.
The goal is to share knowledge that would have been helpful when first starting with Claude Code.

0:57 Foundational Tips: Setup, Shortcuts, and Essential Commands

Run Claude Code in the root directory of the project to ensure proper context zipping.
Use `/init` to analyze the codebase and create an initial `claw.md` file.
The `claw.md` file operates on a hierarchy, with checked-in memory and user memory.

4:15 Structuring the `claw.md` File for Optimal Performance

Aim for a `claw.md` file around 300 lines to avoid bloating the context window.
Include high-level technical architecture, requirements, and domain context in the `claw.md` file.
Incorporate build and validation flows to enable AI self-improvement.

6:19 Essential Keyboard Shortcuts for Efficient Claude Code Usage

`Shift + Tab` toggles between accept edits and plan modes.
Use `Escape` to interrupt Claude Code's processes.
Double-tapping `Escape` clears the input or rewinds to a previous context point.

13:48 Essential Slash Commands for Context Management

`/clear` clears the context, useful for starting new features without old context influencing the new project.
`/context` provides a visual representation of the current context Claude Code is operating in, aiding in auditing and identifying token usage offenders.
`/models` allows switching between models like Opus 4.5, Sonnet, or I coup based on cost sensitivity and workflow requirements.

19:01 Compaction and Model Selection

`/compact` manually compacts the current context window, useful for saving context versions.
`/models` allows selecting different models (e.g., Opus 4.5) based on cost and workflow needs.
`/resume` recovers the old context if an instance is accidentally killed.

21:58 Managing Git and Deep Dive into `claw.md`

Use Git as a safety net for code changes, leveraging Claude Code for summaries and test plans.
The `claw.md` file reads top to bottom, prioritizing rules from top to bottom.
Update `claw.md` rules using Claude Code to avoid manual edits and ensure consistency.

25:50 Advanced `claw.md` Techniques: Keywords and Compound Engineering

Use keywords to trigger specific skills within the `claw.md` file.
Compound engineering involves committing the `claw.md` file into the codebase, improving the AI coding experience for teammates.
Use `claude - dangerously skip permissions` in throwaway environments, but exercise caution with OS-level changes.

34:18 Optimizing Workflows with Plan Mode and Fresh Context

Always start with plan mode to iterate and argue with Claude Code, ensuring a solid context before execution.
Fresh context beats bloated context; spend time building good context initially.
Persist context using a second brain concept, saving and loading contexts for new sessions.

43:27 Leveraging Build Cycles and Model Selection for Efficiency

Focus on the build cycle, using Claude Code to add debug logs, control emulators, and analyze performance traces.
Consider using Opus for optimal performance, especially with multiple Claude instances running in parallel.
Interrupt Claude Code when it makes assumptions or goes off course, correcting it as needed.

Summary, takeaways, and chapters were generated by AI from the video's transcript and may contain errors. The video belongs to its creator, John Kim.