Token Smart Saver Assistant

In local large model dialogues, token consumption is a critical pain point. Each request consumes tokens, and local environments often have limited resources, leading to high costs, slow response times, and constraints from the model's context window. This impacts user experience and practicality.

Problem Background

Specifically, user inputs—whether text, code snippets, or mixed content—often contain redundant information like extra spaces, repeated phrases, or filler words. These redundancies, while not affecting semantics, occupy valuable token quota, causing dialogues to easily exceed model limits or reduce inference efficiency.

How the Skill Works

This Token Smart Saver Assistant addresses these issues through core capabilities:

• Intelligent Content Type Recognition: Automatically identifies input as pure text, code, or mixed content, applying differential compression strategies with high accuracy to avoid semantic loss or code corruption.

• Three Switchable Compression Levels:

• Light Mode: Cleans only formatting issues, preserving original tone.
• Default Mode (Recommended): Removes redundant phrases, simplifies long sentences, balancing effect.

• Strict Mode: Deeply refines sentences, deleting connectors and redundant modifiers for higher compression.

• Text Deep Optimization:

• Intelligent synonym simplification for long sentences, e.g., converting "我想请问一下你能不能帮我写一个爬虫" to "写一个爬虫".
• Automatic deduplication of redundant phrases, filtering meaningless repetitions.

• Normalization of spaces and formats to reduce token waste from formatting.

• Safe Code Compression: Only compresses formatting (e.g., blank lines, indentation), with graded comment cleanup, but does not modify variable names, logic structures, or syntax keywords, ensuring code remains runnable.

• System Command Protection Mechanism: Automatically identifies and skips key commands starting with 系统:, 指令:, etc., preventing functional anomalies.

• Real-time Token Saving Statistics: Displays original tokens, compressed tokens, and saved count per dialogue round for transparency.

• Natural Language Command Control: Users can switch modes with colloquial commands like "严格模式" or "打开省Token提示" without parameter configuration.

Applicable Boundaries and Notes

This skill is suitable for QClaw / OpenClaw local agents in daily dialogues, code queries, long text consultations, etc., especially for long-context dialogues that might exceed model windows. It emphasizes security, no external network dependency, and is ideal for enterprise or personal private deployments.

Notes:
- The skill does not alter code logic or model understanding, only optimizes token consumption.
- Compression strategies are rule-based, which might over-simplify in extreme cases, but default mode balances semantic integrity.
- It runs locally via pure Python execution with no third-party library dependencies, ensuring controllability.