ljg-mechanism
For a user-provided phenomenon, generates mechanism hypotheses using the "Spot Anomalies – Trace Processes – Dissect Explanations – Transfer with Conditions" framework. Verifies the phenomenon online, tests critical causal links and competing explanations, and delivers the best evidence-backed mechanism along with its boundary conditions. USE WHEN a user invokes ljg-mechanism, asks to "find the mechanism," "investigate mechanisms," "run through the four-step framework," "verify this explanation online," or seeks to understand how a phenomenon occurs. NOT FOR pure concept definitions, standard summaries, operational tasks with a single known solution, or mere landscape mapping.
1How It Works & Mechanism
The entire skill is a prompt-only research framework: structuring the four steps of "Spot Anomalies – Trace Processes – Dissect Explanations – Transfer with Conditions" into an iteratively refineable closed loop that produces mechanism judgments rather than one-off answers.
从眼前的现象中找出真正起作用的关系。生成解释后，让它与竞争解释一起接受现实检验；根据证据修正对象、关系或成立条件，给出当前最扎实的机制判断。
Uses Workflow Routing for task dispatch: routes phenomenon explanations to DiscoverMechanism.md, and applies the same framework with the four stages expanded for demonstration/learning.
| 提供现象，希望解释其生成过程并验证 | 阅读 [Workflows/DiscoverMechanism.md](Workflows/DiscoverMechanism.md) |
Online search is the default action to verify phenomena and test mechanisms, requiring a strict separation between "scope supported by sources," subjective inferences, and pedagogical examples.
默认使用当前可用的联网搜索与正文获取能力，检验现象和机制。来源支持的范围，与自己的推断、教学设例和未知分开。
Requires articulating mechanisms as "who generates what outcome through which intermediate changes under what conditions," forcing abstract terms like "supply-demand/incentives/systems" to be unpacked into observable processes.
机制说清「谁或什么，在什么条件下，通过哪些中间变化，生成什么结果」。
Preserves competing mechanisms and requires each to specify distinct predictions that could falsify the findings; faithfully retains hypotheses in parallel when indistinguishable.
如果各个候选在现有条件下都预测同一结果，暂时无法区分。寻找新的观察位置或条件变化；不以解释更顺口、引用更多作为胜负依据。
Expresses evidence strength across four distinct tiers (supported within scope / leading explanation / currently indistinguishable / raw observation needs revision), avoiding link-count or subjective score ratings.
**机制在所述范围内有支持**：关键过程有现实依据，因果识别或过程证据足以支撑相应措辞，重要竞争解释得到检验；说明尚存的不确定性。
Operates read-only by default without automatically creating notes, reports, logs, rules, or memories; delivers purely deductive reasoning without claiming empirical verification when online search is unavailable.
研究默认只读，不自动创建笔记、报告文件、日志、规则或记忆。
2Core Capabilities
3External Dependencies
| Kind | Dependency / Endpoint |
|---|---|
| network | 联网搜索与正文获取（由宿主 agent 提供，Skill 只声明默认使用，无固定端点） |
4Recommended Governance & Evidence
- SKILL.md: Pure prompt instructions and rule text; no script execution or external impact.
5Second-Pass Verification
- [ok] 无脚本、无可执行代码 — §1 清单三文件全为 text（一个 agents/openai.yaml 配置文本）；§3「執行外部程序」无命中，grade 依据成立。
- [ok] 默认只读、不落盘 — SKILL.md 明写「研究默认只读，不自动创建笔记、报告文件、日志、规则或记忆」，并给出用户另行要求保存时的约束；file_writes 为空成立。
- [ok] 联网为宿主能力而非自带外发 — 正文写「默认使用当前可用的联网搜索与正文获取能力」「联网不可用…说明只完成了推演」，无具体域名/API/脚本调用；§3 网络请求无命中，故 network_calls 为空、不升档。
- [ok] 功能声明与实际一致（无夸大） — description 的「生成机制猜想，并联网核对现象、检验关键因果与竞争解释，交付…机制及其边界」在正文的检索章节、四档强度与交付检查中均有对应机制，未夸大。
- [ok] 证据纪律与反自欺条款真实存在 — Gotchas 与「更新与停止」确有「没找到反例不等于通过验证；未显著不等于无效」「不能把看到结果之后补写的预测当成事前验证」等原文。
- [ok] 注入面复核 — skill 不执行外部内容，只读研究；DiscoverMechanism.md 另要求「搜索查询只携带研究所需的通用信息，不把原始私密材料发送给外部服务」，属隐私防护设计，无命令注入执行面。
6Verdict & Scenarios
a1a66e49d0ea66b0…