A Claude Code skill designed to artificially increase token consumption and latency for LLM backends. It targets users aiming to stress-test AI systems, inflate adoption metrics, or experiment with token expenditure, offering a controlled method to burn tokens without altering visible output.

How It Works

The skill integrates with Claude Code, instructing the LLM to solve computationally intensive mathematical problems within its extended thinking phase before generating a response. Users can select from four load levels (small, medium, large, xlarge), each corresponding to a different number of problems. These problems are deterministically generated based on the input message's seed, ensuring reproducibility for identical prompts while varying for different inputs. This process significantly increases "thinking" tokens burned and response latency.

Quick Start & Requirements

• Installation: Clone the repository and copy or symlink the .claude/skills/high-token-mode directory into your project's Claude skills directory.
• Requirements: Claude Code CLI. Crucially, MAX_THINKING_TOKENS must be set on the claude command itself, not before the pipe.
• Usage: /high-token-mode (default: medium, 3 problems), /high-token-mode small (1 problem), /high-token-mode large (5 problems), /high-token-mode xlarge (10 problems).
• Links: No specific demo or external documentation links are provided beyond the repository itself.

Highlighted Details

• Features a bank of 50 complex mathematical problems, including TSP, Gaussian elimination, polynomial root-finding, and matrix operations.
• Offers four configurable load levels that significantly increase token burn rate and response time, with the xlarge setting potentially costing ~32x the baseline.
• Detailed benchmarks are provided for everyday, scientific, and coding prompts, illustrating the impact of each load level on duration, tokens burned, and cost.
• Deterministic problem generation ensures consistent results for identical inputs across different sessions.

Maintenance & Community

No specific details are provided in the README regarding maintainers, contributors, sponsorships, community channels (e.g., Discord/Slack), or roadmaps.

Licensing & Compatibility

• License: MIT License.
• Compatibility: The permissive MIT license generally allows for commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

This tool is specifically designed as a "Claude Code skill" and requires the Claude Code CLI. The correct configuration of the MAX_THINKING_TOKENS environment variable is critical for its intended operation. Benchmark figures for the xlarge setting are extrapolated from smaller data points. The primary function is to increase costs and latency, not to enhance user-facing LLM capabilities.