SkCC is a compiler for LLM agents. Author skills once as SKILL.md, compile to Claude Code, Codex, Gemini CLI & Kimi CLI — with built-in security hardening.
LLM agents are highly sensitive to prompt formatting — the same SKILL.md can vary up to 40% in performance across frameworks. Manual per-platform rewriting is unsustainable.
Inspired by classical compiler design (LLVM, MLIR). A unified SkIR decouples skill semantics from framework-specific formatting.
Compile to Claude Code (XML), Codex (XML-Markdown), Gemini CLI (Markdown+YAML), and Kimi CLI (Full Markdown) from a single source.
Automatic compile-time safety constraint injection. 4 rule categories covering HTTP, Loop, DB, and Parse safety — 94.8% trigger rate on real skills.
Rust-native, zero-copy parsing with Arc
Routing manifest generation (~50 tokens/skill) for efficient agent initialization. Full skill content loaded only on semantic match.
Low → Medium → High → Critical classification with graduated enforcement. HITL triggers for high-risk operations.
10–46% runtime token reduction and 23–43% execution time reduction across frameworks. Structured formats reduce trial-and-error.