SkillPilot: Innovative Practice of Agent Skill Router

SkillPilot is a universal agent skill router. By performing skill routing before LLM inference, it addresses three major pain points in current AI assistant skill selection: slow speed, high cost, and low accuracy. It uses vector semantic matching technology to reduce routing time from 1-5 seconds to under 25 milliseconds, significantly improving performance and efficiency.

Modern AI assistant frameworks (such as OpenClaw, Claude Code, Codex, etc.) face skill selection challenges as the number of skills increases. The mainstream approach is to put all skill descriptions into system prompts for LLM decision-making, which brings three major problems:

• Speed Issue: Each request requires waiting 1-5 seconds for LLM inference, which is unacceptable in terms of latency;
• Cost Issue: Skill descriptions containing thousands of tokens result in high API fees that rise linearly with the growth of the skill library;
• Accuracy Issue: The more skills there are, the easier it is to get confused. When dealing with similar skills, it's hard to choose accurately, leading to wrong operations or irrelevant responses.

The core innovation of SkillPilot lies in changing the timing and method of skill selection: instead of relying on LLM decisions, it quickly selects skills through vector semantic matching before LLM inference. The process is as follows: User query → Fast path (<2ms) → Semantic matching (vector similarity <20ms) → Conflict resolution (<5ms) → Execute skill or inject context. The total routing time is less than 25 milliseconds, which is a performance improvement of over 200 times.

SkillPilot's architecture is modular and scalable, with core components including:

• Platform Adaptation Layer: Supports OpenClaw plugins (npm package), Claude Code hooks, LangChain tools, and CLI interfaces, enabling seamless integration with different AI frameworks;
• Core Routing Engine: Fast path (keyword/trigger phrase matching, <2ms), semantic path (vector similarity matching, <20ms, based on SQLite vector database), conflict resolution (priority rules + context, <5ms);
• Skill Index System: Stored in SQLite, automatically parses SKILL.md to extract semantic embedding vectors, intent patterns, keywords, etc. New skills can be routed with zero manual configuration.

SkillPilot features:

• Zero Configuration: Automatically extracts skill fingerprints without manual configuration; supports skill authors to add routing configurations (trigger phrases, priority, etc.) in SKILL.md metadata;
• Self-Learning: Records routing feedback and automatically adjusts weights to optimize accuracy;
• Conflict Resolution: Identifies similar skill groups (e.g., github series) and prompts users to add prefer_when configuration to eliminate ambiguity.

SkillPilot's performance on a test dataset of 50 intents ×20 skills:

• Accuracy: 93.0%;
• P50 latency: 12 milliseconds;
• P99 latency:23 milliseconds. This proves that it outperforms traditional LLM-based routing in both speed and accuracy.

Ecosystem Integration: Supports multiple deployment methods (global installation, Node.js dependency, source code build, npx run) and shared skill index (multiple platforms use the same library); Application Scenarios:

• Enterprise AI assistants (integrate dozens or hundreds of internal tools to reduce costs and improve efficiency);
• Skill market (third-party developers can easily add skills);
• Real-time scenarios (voice assistants, collaboration tools with millisecond-level latency);
• Edge devices/API-restricted environments (reduce token consumption and lower costs).

SkillPilot represents an important direction for AI assistant architecture: separating decision logic from expensive LLM calls and completing pre-filtering through efficient methods. The layered architecture improves performance, maintainability, and scalability. As AI assistant capabilities expand, skill routing will become key infrastructure, and SkillPilot's open-source implementation provides reference and practice for this field.