Skip to content

overview.md

Latest commit

 

History

History
34 lines (27 loc) · 1.55 KB

overview.md

File metadata and controls

34 lines (27 loc) · 1.55 KB
Raw

Overview

SkillEngine is a framework-agnostic execution engine for LLM agent skills — units of capability defined as Markdown files with YAML frontmatter.

The pipeline

Skill files (Markdown + YAML)
        │
        ▼
   ┌─────────┐    ┌─────────┐    ┌─────────┐    ┌─────────┐
   │ Loader  │───▶│ Filter  │───▶│ Runtime │───▶│ Adapter │
   └─────────┘    └─────────┘    └─────────┘    └─────────┘
   parse YAML    eligibility    shell / VM     LLM tool calls
                 (bins, env)    / code mode
  • Loader — turns a SKILL.md into a Skill dataclass.
  • Filter — decides whether the skill is eligible right now (required binaries, env vars, OS, custom predicates).
  • Runtime — executes commands from skill bodies (bash, code-mode, sandboxed).
  • Adapter — exposes skills as tools for OpenAI / Anthropic / OpenAI-compatible LLMs and translates tool-call results back.

Why progressive disclosure

The system prompt contains only skill names and descriptions. The full body of a skill is only loaded when the LLM calls the skill tool to ask for it. This is the same pattern Claude Agent Skills uses, and it keeps the context window cheap even with hundreds of skills installed.

See skills.md for the file format and agents.md for how AgentRunner orchestrates this.