What should I put in my CLAUDE.md file?

Include only critical information Claude cannot infer, such as your tech stack, coding conventions, domain knowledge, and specific project rules. Keep the entire file under 200 lines.

How do I manage context when responses start to degrade?

Use the /compact command to compress the conversation context. You should also store persistent knowledge in a MEMORY.md file index to ensure information is retained across sessions.

Why should I use plan mode for complex tasks?

Plan mode forces Claude to outline its approach before generating code. This prevents it from writing incorrect code based on incomplete context.

What are subagents used for in Claude Code?

Subagents handle verbose tasks like capturing logs or running code reviews. This splits exploration from editing and keeps your main context window clean for higher-level reasoning.

Claude Code Best Practices for Context and Workflows

Treat Claude as a capable but context-limited junior developer by managing its context carefully and planning tasks before implementation.

Keep your CLAUDE.md files short, ideally under 200 lines, focusing only on tech stacks, conventions, and domain rules Claude cannot infer on its own. Use MEMORY.md files and the /compact command to retain knowledge across sessions and compress context when responses degrade.

Always use plan mode for complex tasks, break work into small incremental steps, and let Claude fix its own bugs to improve its understanding of your codebase. Use subagents to keep the main context window clean, and route mechanical tasks to cheaper models to save tokens.

Claude Code Best Practices for Context and Workflows — infographic

To maximize your productivity and avoid common pitfalls with Claude Code, focus on meticulous context management, iterative planning, and leveraging Claude's advanced features like CLAUDE.md files and subagents. Users emphasize that treating Claude as a highly capable but context-limited junior developer will yield the best results.

Optimize Context Management

Keep CLAUDE.md Concise and Strategic: Create CLAUDE.md files in your project root and within specific directories to provide Claude with essential, up-to-date project context. "Keep CLAUDE.md short."

Utilize CLAUDE.md for Core Information: Focus on including only critical information like tech stack, coding conventions, domain knowledge, and specific project rules that Claude couldn't infer on its own. "Keep it under 200 lines (The Golden Rule)."

Manage Context Throughout Sessions: Use /compact to compress conversation context when responses degrade, and store persistent knowledge, such as infrastructure state and lessons learned, in MEMORY.md to ensure Claude retains crucial information across sessions. "The memory system most people skip: /home/user/.claude/projects/<project>/memory/ with a MEMORY.md index + topic files."

Adopt Iterative Workflows

Emphasize Planning Before Implementation: Always begin complex tasks by having Claude outline its approach in "plan" mode (Shift+Tab twice) to prevent generating incorrect code based on incomplete context. "Use 'plan' mode for anything complex."

Break Down Tasks into Small Iterations: Instead of requesting a full feature implementation, guide Claude through small, incremental steps, reviewing each one before proceeding to the next. "Only ask for the first step."

Let Claude Self-Correct and Verify: Allow Claude to fix its own bugs to improve its understanding of the codebase and explicitly instruct it to verify plans against existing architectural patterns. "Don't fix bugs yourself, let Claude fix them."

Leverage Advanced Features and Tools

Implement Skills and Hooks for Automation: Create custom skills for repetitive workflows and use hooks (.claude/settings.json) to automate actions, enforce rules, or manage context programmatically. "Hooks are the real power feature."

Utilize Subagents for Focused Tasks: Employ subagents to handle verbose tasks like capturing logs or running code reviews, keeping your main context window clean for higher-level reasoning. "Subagents split exploration from editing."

Integrate External Tools and Other Models: Connect Claude to internal tools, data sources, and APIs via MCP servers, and consider routing mechanical tasks through cheaper LLMs to save tokens and optimize resource usage. "Stop sending every file read and git status to opus."

Are you currently using any of these best practices in your Claude Code workflow?

Best for: Developers who want to maximize productivity and avoid common pitfalls when using Claude Code.