Fixing Claude Code's Token Bloat: 13 Practical Optimization Tips

EXM7777 · x · 2026-08-11

The author points out that Claude Code's default configuration is bloated, consuming excessive tokens and suffering from slow, argumentative responses. They share a comprehensive optimization workflow centered around the global CLAUDE.md file and system configurations.

Key Optimization Tips:

Streamline System Prompts: Instruct the model in CLAUDE.md to use ASD-STE100 simplified technical English and report only decision-critical elements. Keep the file under 200 lines and rebuild it from scratch every 2-3 months.
Audit Context & Usage: Use /context to inspect what fills the context window before typing, and /usage to identify which MCP server, skill, or plugin is burning tokens.
Prefer CLIs over MCPs: CLI commands return only the requested lines, whereas MCPs tend to dump their entire result into the context window.
Disable Idle Components: Turn off unused MCP servers and set rarely used skills to manual-only (disable-model-invocation: true) to save tokens.
Permissions & Effort: Write restrictive rules as permission settings rather than prompt sentences; restrict high effort to planning phases and use medium/low for execution.
Session Management: Run /clear at checkpoints or when exceeding 400k context. Offload vision tasks to headless workers and heavy exploration to subagents, ensuring the main window only processes summaries.

Original post →

Fixing Claude Code's Token Bloat: 13 Practical Optimization Tips

More from coding & agent