author-skill

原文涵盖的主题

• Author a Reusable, Opinionated Agent Skill (author-skill)
• 0. The Five Laws of Modern SOTA Skill Engineering (現代 SOTA Skill 編寫五大鐵律)
• 1. Methodology vs. Organization State & Four-Layer Guidance Architecture
• Auto-Triggered Skill Architecture & Daily Mistake-to-Skill Loop
• General, reusable and future-proof skill evolution
• 2. Stage 1: Define the Scenario Trigger and Positive/Negative Cases
• 3. Stage 2: Structure SKILL.md and Colocated references/
• 4. Stage 3: Verify Links, Format, and Trigger Discrimination

Apply the documentation and plain-language standards in references/docs.md and ../clear-writing/SKILL.md. A skill is asynchronous Agent-to-Agent (A2A) communication across time: a past specialist sharing hard-won domain reality with a future frontier SOTA model. Just as live A2A messages between peers must never issue imperative orders or pre-chew solutions that stop the other agent from thinking, a skill must never over-constrain a SOTA model with rigid scripts, checklists, or micro-rules that hijack its first-principles reasoning—nor leave it with zero boundaries where default model priors (such as RLHF people-pleasing or HTTP 200 OK completion illusions) cause preventable harm. Reusable guidance obeys mechanism, domain-boundary and causal invariants: keep organization policy in caller configuration, use role-neutral examples and synthetic policy fixtures, and express exact causal preconditions rather than identity guesses or fail-open defaults. Before adding or editing any line in a skill, test it against all five laws: | Law | ❌ Over-Constrained (過分約束 / 命令式預嚼 — Kills SOTA Thinking) | ❌ Under-Constrained (完全無約束 / 空泛口號 — Zero Protection) | ✅ SOTA Skill Sweet Spot (鎖邊界、給變量、傳現象、解放思

这里只展示节选，请以完整原文和当前要求为准。