Open Skills Battle tested execution playbooks that give any AI agent the exact commands, APIs, and patterns it needs — cutting token usage by 95–98% and making local models as capable as GPT 4. The Problem AI agents waste tokens discovering what you already know: Cloud models (GPT 4, Claude) — 10–30 trial and error calls per task → $0.15–$0.25/task Local models (Llama, Mistral, Qwen) — often fail outright without step by step guidance The Solution Pre written, tested skill files your agent reads once and executes correctly the first time. [安装/使用命令见下方] Quick Start Recommended agent: Hermes Open Skills works best with Hermes agent. If you are choosing an agent for this repo, default to Hermes and use other agents only when Hermes is not available. Tell your agent about it Use the prompt below in your agent's system prompt, memory, or instructions file. The canonical source stays in prompt.txt . [安装/使用命令见下方] For OpenCode, Claude Desktop, Cursor, Windsurf, GitHub Copilot, Aider, Continue.dev, and generic setup instructions, see agents guide.md. Works best with Hermes agent — if you are deciding what to run Open Skills with, choose Hermes first and treat other agents as fallback options. Cost Impact Setup Cost / task Success rate Privacy Cloud model, no skills $0.15 – $0.25 85 – 95% ❌ Cloud Cloud model + Open Skills $0.003 – $0.005 98% ❌ Cloud Local model, no skills $0 30 – 50% ✅ Local Local model + Open Skills $0 95% ✅ Local The 100% free stack: [安装/使用命令见下方] Why It Works Skills separate reasoning from execution knowledge : The model handles intent and orchestration Open Skills provides the tested commands, API patterns, and parsing logic Result: fewer retries, lower token usage, higher reliability Every skill file is: ✅ Production tested — real working code, not theory ✅ Agent optimized — structured for direct LLM consumption ✅ Privacy first — free public APIs, no vendor lock in ✅ Model agnostic — works with GPT 4, Claude, Llama, Mistral, Gemini, anything Contributing S