Universal Claude.md – cut Claude output tokens

One file. Drop it in your project. Cuts Claude output verbosity by ~63%. No code changes required. Note: most Claude costs come from input tokens, not output. This file targets output behavior - sycophancy, verbosity, formatting noise. It won't fix your biggest bill but it will fix your most annoying responses. Model support: benchmarks were run on Claude only. The rules are model-agnostic and should work on any model that reads context - but results on local models like llama.cpp, Mistral, or others are untested. Community results welcome.

When you use Claude Code, every word Claude generates costs tokens. Most people never control how Claude responds - they just get whatever the model decides to output.

By default, Claude:

• Opens every response with "Sure!", "Great question!", "Absolutely!"
• Ends with "I hope this helps! Let me know if you need anything!"
• Uses em dashes (--), smart quotes, Unicode characters that break parsers
• Restates your question before answering it
• Adds unsolicited suggestions beyond what you asked
• Over-engineers code with abstractions you never requested
• Agrees with incorrect statements ("You're absolutely right!")

All of this wastes tokens. None of it adds value.

Drop CLAUDE.md

into your project root. Claude Code reads it automatically. Behavior changes immediately.

your-project/
└── CLAUDE.md <- one file, zero setup, no code changes

This file works best for:

• Automation pipelines with high output volume (resume bots, agent loops, code generation)
• Repeated structured tasks where Claude's default verbosity compounds across hundreds of calls
• Teams who need consistent, parseable output format across sessions

This file is not worth it for:

• Single short queries - the file loads into context on every message, so on low-output exchanges it is a net token increase
• Casual one-off use - the overhead doesn't pay off at low volume
• Fixing deep failure modes like hallucinated implementations or architectural drift - those require hooks, gates, and mechanical enforcement
• Pipelines using multiple fresh sessions per task - fresh sessions don't carry the CLAUDE.md overhead benefit the same way persistent sessions do
• Parser reliability at scale - if you need guaranteed parseable output, use structured outputs (JSON mode, tool use with schemas) built into the API - that is a more robust solution than prompt-based formatting rules
• Exploratory or architectural work where debate, pushback,