45 Most Worth-Knowing Tips for Claude Code After a Year

↓ The conclusion first: most people only use about 30% of its potential.

───

🔋 Save Tokens / Reduce Costs

Claude Code's default system prompt consumes nearly 20k tokens—burning 10% of your context before you even start working. After patching, it can be reduced to around 9k, saving about 10k tokens per conversation. The difference is significant with long-term use.

For MCP tools, it's recommended to enable lazy loading (ENABLE_TOOL_SEARCH: "true"). Tools are only loaded when needed, not occupying space unnecessarily.

Actively use /compact or write a HANDOFF.md to hand over to a new session. This is much more precise than waiting for automatic compression.

───

🧠 Context Management

AI context is like milk—the fresher, the better. Start a new session for new topics; don't force continuation.

For ultra-long conversations, don't rely on summaries. Use "semi-cloning": keep only the real messages from the latter half of the conversation. This preserves information without distortion. You can configure a hook to trigger automatically when context exceeds 85%.

───

🚀 Workflow Speed Boost

Voice is faster than typing, and significantly so. Even if local Whisper models occasionally make transcription errors, Claude can still understand. You can even use it on a plane.

For multitasking, the "cascade method" is recommended: open 3-4 terminal tabs side-by-side and process tasks from left to right in rotation. This keeps you focused on one direction without jumping around.

Git worktree: Work on multiple branches in separate directories independently without interference. Combined with multiple tabs, the efficiency doubles.

Claude can't access sites like Reddit? Configure Gemini CLI as a fallback, relay through tmux, and research tasks can still run.

───

🔍 Code Quality

TDD pairs surprisingly well with Claude Code: First, write and commit a failing test, then let Claude write the code to pass the test. With a clear goal, its accuracy is much higher.

Don't sift through GitHub Actions logs yourself. Use /gha <url> for one-click analysis: It automatically determines if it's a sporadic flake, a breaking commit, or a configuration issue, providing precise diagnosis.

Regularly run cc-safe to scan your settings.json, checking if you've accidentally approved dangerous commands like rm -rf or sudo.

───

🤖 Advanced: Multi-Agent Orchestration

Put Claude Code into a Docker container, start it with --dangerously-skip-permissions, and let it iterate, fix, and test autonomously. If something goes wrong, it's contained within the sandbox, keeping the host safe.

Go further: Have an outer Claude control a worker Claude inside the container via tmux. Send commands, read outputs, pull results—all without manually switching terminals. A fully automated agent pipeline.

Need to analyze a large codebase in parallel? Let Claude launch multiple sub-agents simultaneously, dividing work across different modules. You can even allocate compute power by task complexity, specifying Opus / Sonnet / Haiku models.

───

📁 Configuration Habits

Start CLAUDE.md from an empty file. Only add things you find yourself saying repeatedly. Keep it lean. Review it periodically and delete outdated content.

Set up terminal aliases: c = claude, ch = claude --chrome. The time saved daily is no joke.

Disable auto-updates (DISABLE_AUTOUPDATER: "1") to prevent new versions from breaking your system prompt patches.

───

The complete list of 45 tips + a one-click installation script (configuring status bar, dx plugins, aliases, lazy-loaded MCP all at once):

👉 https://t.co/28rVNYEQvR