Title: SkCC: Enabling Portable and Secure Skill Compilation for Cross-Framework LLM Agents

Abstract:

While Large Language Model (LLM) agents are increasingly dependent on reusable skills, such as $SKILL.md$ files, to handle complex operations, these components suffer from a significant lack of portability. Because agent frameworks are highly sensitive to prompt formatting, the same skill can yield widely varying performance results depending on the platform. Despite this, skills are typically written once in a format-agnostic Markdown style, which forces developers to incur the high cost of rewriting them for each specific framework. Furthermore, this approach largely neglects security, resulting in widespread practical vulnerabilities.

To resolve these issues, we introduce SkCC, a compiler designed for LLM agents that integrates classical compilation principles into skill development. The core of SkCC is SkIR, a strongly-typed intermediate representation that separates skill semantics from framework-specific formatting requirements, thereby facilitating portable deployment across diverse agent ecosystems. Built upon this IR, a static Optimizer enforces security constraints to block potential vulnerabilities prior to deployment.

SkCC operates as a four-phase pipeline that significantly reduces adaptation complexity. Specifically, it lowers the computational burden from $O(m \times n)$ to $O(m + n)$, where $m$ represents the number of skills and $n$ denotes the number of frameworks. Experimental results on SkillsBench indicate that SkCC provides consistent and substantial improvements over original methods. Performance gains include a pass rate increase from 21.1% to 33.3% on Claude Code and from 35.1% to 48.7% on Kimi CLI. Additionally, the design ensures sub-10ms compilation latency, a 94.8% proactive security trigger rate, and runtime token savings ranging from 10% to 46% across various frameworks.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC