Always-on token-efficiency skill for coding agents like Claude Code, Codex, and Cursor. It addresses high token consumption and associated costs by trimming output across sessions, yielding significant cost reduction (~31% average bill savings) without sacrificing correctness, enabling more economical AI assistant use.

How It Works

The skill injects a session-start directive via agent-specific hooks or context files. It systematically reduces tokens by removing conversational filler, minimizing documentation verbosity, optimizing test selection, enforcing concise code, and improving context management. Configurable telegraphic levels (ultra, lite, on, off) allow users to balance conciseness with necessary detail.

Quick Start & Requirements

Install via curl -fsSL https://raw.githubusercontent.com/Kulaxyz/token-diet/main/install.sh | bash. Options like --ultra are passed via bash -s --. Target agents with -a <agent> or install project-locally with --project. On-demand control: /token-diet [on|lite|ultra|off]. No specific prerequisites beyond a compatible agent environment are listed.

Highlighted Details

• Average bill reduction of ~31% (-17% to -54% by session type).
• Output savings of -30% to -81% in Sonnet 5 runs, with correctness maintained.
• Applies specific trimming for replies, documentation, code, context, and tool usage.
• Configurable levels: on (default), lite (communication/artifacts), ultra (telegraphic chat).

Maintenance & Community

The README provides no details on contributors, sponsorships, or community channels.

Licensing & Compatibility

No software license is specified in the README. This requires further investigation for commercial use or integration.

Limitations & Caveats

Savings vary by session type (output-heavy vs. read-heavy). The "always-on" functionality depends on agent internals that may change. The absence of a stated license is a significant adoption blocker.