Skill Writer

Language: English · 简体中文

A cross-platform meta-skill for creating, evaluating, and optimizing AI assistant skills through natural language interaction.

What is a skill?

A skill is a small instruction file (Markdown, .md) that you place in your AI assistant’s skills folder. It tells the assistant how to handle a specific type of request — like “summarize a git diff”, “write a PR description”, or “validate API responses”. When you say something that matches the skill’s trigger phrases, the assistant follows the skill’s instructions automatically.

Think of it like a custom command: you define what it does once, and then the assistant does it consistently every time you ask.

After creating a skill, it lives in ~/.claude/skills/ (or your platform’s equivalent). Restart the assistant to activate it.

---

Overview

Skill Writer is a meta-skill that enables AI assistants to create, evaluate, and optimize other skills through natural language interaction. No CLI commands required - just describe what you need.

Key Features

• Agent Install: One-line install via “read [URL] and install” — works in any supported platform
• Zero CLI Interface: Natural language interaction - no commands to memorize
• Cross-Platform: Works on 8 platforms — Claude, OpenClaw, OpenCode, Cursor, Gemini, OpenAI, Kimi, Hermes
• Nine Powerful Modes: CREATE, LEAN, EVALUATE, OPTIMIZE, BENCHMARK, INSTALL, COLLECT, SHARE, and GRAPH
• Template-Based: 4 built-in templates for common skill patterns
• Quality Assurance: 1000-point scoring system with certification tiers
• Tier-Aware Evaluation: Tier-adjusted scoring weights for planning / functional / atomic skills (three-tier skill hierarchy)
• Reliable LEAN Scoring: 16 checks split into [STATIC] (deterministic, 335 pts, zero variance) and [HEURISTIC] (LLM-judged, 165 pts) — score variance documented per phase
• Security Built-In: CWE-based + OWASP Agentic Skills Top 10 (ASI01–ASI10) detection + supply-chain trust verification for pulled skills
• Continuous Improvement: Automated optimization with convergence detection + co-evolutionary VERIFY step + persistent score history
• Self-Evolution: UTE (Use-to-Evolve) L1 in-session always active; L2 collective with hooks/backend
• Empirical A/B Benchmarking: BENCHMARK mode runs parallel API calls (skill vs. baseline), blind-grades via Comparator agent, reports delta_pass_rate + token overhead + PASS/FAIL verdict (v3.5.0)
• Honest Skill Labeling: generation_method + validation_status fields prevent unvalidated skills from silently reaching production
• Behavioral Verifier: Optional task-sample testing produces a pragmatic_success_rate independent of theoretical score