A compact Codex skill for keeping engineering lean and academic writing centered on the strongest supported contribution.
- Solve the stated problem with the smallest general change; reuse existing implementations before rebuilding.
- Avoid speculative checks, silent fallbacks, test-specific patches, unnecessary abstractions, and scope creep.
- Treat normal GPU and numerical variation with tolerances and claim-relevant evidence.
- Write papers as claim-centered releases; give every experiment an argumentative role and state decisive advantages in prose.
- Remove self-weakening language, imagined reviewer objections, repetitive contrast templates, literal translations, and AI-styled prose.
- Keep necessary limitations exact without letting them dominate the contribution.
SKILL.md contains the core rules and routes each failure mode to one short file under cases/. Read only the relevant case. Cases cover ML nondeterminism, test patching, test overengineering, reuse, scope creep, defensive paper writing, the release principle, and academic terminology.
Place this directory in a Codex skill search path, or point Codex to its SKILL.md. Invoke it by name:
Use $avoid-overkill to review this code or paper and keep it proportional, evidence-grounded, and direct.
这是一个供 Codex 使用的短 Skill，约束两类倾向：工程上无依据地增加机制，论文中把最强贡献写成工作汇报或自我辩护。
- 工程克制：围绕明确问题做最小且可泛化的改动，不添加推测性校验、静默降级、测试补丁、无谓抽象和相邻改动。
- 证据相关：用容差和会影响结论的证据处理 GPU 或数值波动，不追求无意义的逐字节一致。
- 优先复用：先查仓库、已有依赖、标准库和官方实现，再决定是否自建。
- 发布会原则：论文围绕最强且有证据支持的优势组织，不按尝试和失败的时间线写成工作汇报。
- 实验有职责：每个实验都应支撑主效应、机制、边界、稳健性或必要权衡；没有论证作用就删。
- 表达主动：正文明确说出关键优势，删除自我削弱、预写审稿意见、固定对比句式、生硬直译和 AI 腔。
SKILL.md 仅保留核心规则，并将具体场景分散到 cases/。使用时只读相关案例。案例涵盖机器学习非确定性、测试补丁、测试过度工程、代码复用、范围蔓延、防御性论文写作、发布会原则和专业术语。
1. 本地安装
克隆仓库并将其添加至 Codex 或 Coding Agent 的 Skill 搜索路径：
git clone https://github.com/yeahjack/avoid-overkill2. 远程同步
直接在对话中让 Coding Agent 安装该 Skill：
请安装 https://github.com/yeahjack/avoid-overkill 中的 Skill。