Using Claude Code as a Daily Tool: Claude.md, Skills, Subagents, Plugins, MCP

Claude Code's productivity goes beyond simple prompt writing and is determined by how you accumulate and verify memory, custom commands, parallel sessions, and project settings. The key is building an environment where Claude can verify its own results.

Hierarchical Configuration Using the .claude Directory The .claude/ directory functions as a hierarchical configuration system that separates project and global scopes.

• Project scope: Configured in .claude/ within the repository and committed to Git to share with the team.
• Global scope: Located at ~/.claude/, containing personal preferences applied across all projects on the local machine.

Key Components and Roles

• CLAUDE.md: An instruction file loaded every session. It should be kept short and clear, and it's best to focus on build commands and verification sequences.
• Skills: Reusable specialized prompts that can be invoked with /name.
• Subagents: Agents that perform reviews, debugging, and migrations in a separate context.
• Rules: Allow defining topic-specific instructions that apply only to certain paths.
• MCP (Model Context Protocol): Configuration for external tools and data connections shared by the project team.

Efficient Workflow: Explore → Plan → Implement It's important to treat Claude not as a mere pair programmer but as a delegated engineer.

• Using Plan Mode: Enter plan mode via Shift+Tab to write a design document and go through a review process before implementation.
• Verification-Centered Operation: When Claude makes a mistake, instruct it with "Update CLAUDE.md so you do not repeat this" to automatically accumulate rules that prevent mistakes.
• Precise References: Instead of vague instructions, it's effective to directly specify files like @file_path or pass error logs through pipes.