Kill test first
三个实战验证的Claude Code Skill：会话交接、CLAUDE.md 生成、事前否决审计
npx -y skills add redamancy231-create/claude-skills --skill kill-test-firstAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 26 days oldThe repository was created 26 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
[决策:预注册] Pre-registration audit for new research directions, quant strategies, or thesis topics: falsifiable hypothesis, prior-art check, strongest counter-argument, death criteria before results, cheapest falsification with strong baseline, three-state verdict (GO/STOP/REDESIGN). Use before committing resources to a new direction. Not for bug fixes, experiment reproduction, or routine controlled experiments.
The file declares its own license as CC-BY-4.0. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
4.0 KB, as published. Nobody here has run it
Kill-Test First（事前否决 + 预登记）
核心信念：最贵的错误不在执行里，而在「根本该不该做」这个没人事前问的问题上。 真正 alpha：不是「能发现失败方向」，而是把通用好习惯——强基线 + 预登记 + 反方论证——强制前置，并留下审计痕迹。 强度：你是 decision owner，最终 GO/STOP 你拍板。但独立 reviewer 的建议必须单独冻结；要 override 记一条
override_reason。
快速参考（路由表）
| 场景 | 执行 |
|---|---|
| 新研究想法/策略/选题 → 「值不值得做？」 | 门 1 完整协议（步骤 0→8）→ gate1-protocol.md |
| 已批准项目内的常规实验 | 轻量提醒：仅确认可证伪命题（步骤 1） |
| GO 之后、正式验证数据之前 | 门 2 预登记冻结 → gate2-preregistration.md |
| 量化策略（含回测/因子） | → 额外过「量化策略红线」→ quant-redlines.md |
| 协议概念/合格标准速查 | → minimum-standards.md |
| 案例演示 | → EXAMPLES.md |
触发分级
应触发（走完整协议）：新研究想法 / 新量化策略 / 新论文选题 / 新模型评估 / 新架构方案 / 新数据管线；问「值不值得做 / 有没有 alpha」；准备进入 Phase 1 建设。
不应触发（负例，别强制）：修 bug、复现实验、补测试、已冻结协议内的对照臂、已批准项目内的常规实验。
判定问句：「这是 Phase 1 建设前的新方向决策，还是已批准项目内的常规实验？」前者走完整协议；后者只轻量提醒＝仅确认可证伪命题（步骤 1）+ 问一句「要不要走完整 kill-test」，不强制全协议。
用户直接说「帮我实现 / 帮我做个 X」且 X 是新方向时，也先轻量问一句「要不要先过一道 kill-test？」——别因为没问「值不值得」就跳过。
协议路由
新方向决策 → 门1 事前否决（步骤 0→8）
├─ STOP → 不进 Phase 1，不开工
├─ REDESIGN → 重写命题（累计 ≤2 次，超过升级为 STOP）
└─ GO → 门2 预登记冻结 → 进入下一阶段（死亡判据全程挂着）
已批准常规实验 → 轻量提醒（仅步骤 1，不强制全协议）
参考文件
| 文件 | 内容 |
|---|---|
| reference/gate1-protocol.md | 门1 事前否决完整协议（8步 + 固定输出块） |
| reference/gate2-preregistration.md | 门2 预登记冻结 |
| reference/quant-redlines.md | 量化策略 6 条不可外包红线 |
| reference/minimum-standards.md | 关键概念最低合格标准 |
| reference/death-criteria.md | 本 Skill 的死亡判据 |
| EXAMPLES.md | 实战案例（事后回放） |
审计痕迹说明：「留审计痕迹」指固定输出块中的结构化字段（Independence Declaration / Prior-Art Check / Result / Decision 等），留存于对话记录中。本 Skill 不写文件。
扩展资源：reference/complete-pipeline.md（非执行必需，日常使用可忽略）。