Anything2Skill: Compiling External Knowledge into Reusable Skills for Agents

Researchers introduce Anything2Skill, a framework that converts external knowledge sources into reusable, executable skills for AI agents. By combining skill extraction with retrieval-augmented generation, the system achieves 98.85% success on command-line tasks and 94.10% on GitHub operations, significantly outperforming RAG-only approaches.

Anything2Skill addresses a critical limitation in current AI agent architectures: while retrieval-augmented generation (RAG) excels at surfacing relevant information, agents must still repeatedly infer procedures from raw data. This framework fundamentally shifts the paradigm by pre-compiling procedural knowledge into structured, reusable skills stored in a SkillBank.

The approach tackles a real problem in AI development. Traditional RAG systems retrieve fragmented evidence—code snippets, documentation, logs—forcing agents to redundantly decompose this raw material into actionable procedures. Anything2Skill inverts this workflow by extracting and organizing skills upfront, complete with invocation conditions, contraindications, workflow steps, and confidence scores. This mirrors how humans build expertise: not through repeated re-reading of manuals, but through internalized procedural knowledge.

The technical implementation demonstrates sophistication through taxonomy-guided compilation, registry-level reconciliation, and versioned lifecycle tracking. The system manages skill currency and prevents conflicts while enabling agents to retrieve both declarative evidence and procedural guidance simultaneously. Experimental results are compelling: 98.85% success on qsv (command-line tool operations) and 94.10% on GitHub-CLI tasks substantially exceeds RAG-only baselines.

This advancement has implications for AI agent reliability and scalability. As autonomous systems tackle increasingly complex domains, the ability to consolidate domain expertise into reusable, version-controlled skills becomes operationally critical. The framework's persistent memory approach suggests a path toward agents that accumulate and leverage institutional knowledge more efficiently than current implementations.

• →Anything2Skill compiles external knowledge into structured procedural skills, dramatically improving agent task success rates from RAG-only baselines.
• →The framework extracts skills through plan-and-expand decomposition and manages them via a persistent SkillBank with versioning and taxonomy awareness.
• →Combined approach achieved 98.85% success on CLI tool operations and 94.10% on GitHub tasks, substantially outperforming traditional RAG systems.
• →Skill contracts specify invocation conditions, constraints, evidence, and confidence scores, enabling explicit capability reuse rather than repeated inference.
• →The methodology suggests a scalable architecture for building agents that accumulate and leverage domain expertise through compiled procedural memory.