中文说明：这个 Skill 是为了减少 GPT-5.6 模型（及同类 Coding Agent）的防御性编程问题 —— 为防御想象出来的失败而添加无依据的校验、回退、重试、兼容路径、抽象和范围蔓延，以及在论文写作中产出自我削弱、防御性的表述。当模型可能过度防御、过度设计时，用本 Skill 让改动与行文始终与结果和证据成比例。
About: A compact skill that keeps engineering and academic writing proportional to the outcome and the evidence — the smallest general solution, claim-centered prose. Built to counter defensive programming in GPT-5.6 and similar coding agents: speculative checks, silent fallbacks, retries, compatibility paths, abstraction sprawl, scope creep, and self-weakening paper prose.
Release v0.1.0 · English · 中文
- Engineering: solve the stated problem with the smallest coherent change; no speculative checks, silent fallbacks, test patching, unrequested refactors, or scope creep.
- Evidence: treat normal ML/GPU variation with tolerances and claim-relevant evidence instead of byte-exact reproducibility chases.
- Reuse: prefer repository, dependency, standard-library, or official implementations; custom-build only the unsupported part.
- Paper writing: treat the manuscript as a release, not a work log — organize around the strongest supported claim, give every experiment one argumentative role, and remove self-weakening or defensive prose.
This repository includes a Codex plugin manifest at .codex-plugin/plugin.json and a Codex marketplace at .agents/plugins/marketplace.json.
codex plugin marketplace add 6Kmfi6HP/avoid-overkill
codex plugin add avoid-overkill@avoid-overkillYou can also install it interactively from the Avoid Overkill marketplace tab in /plugins.
This repository includes a Claude Code plugin manifest at .claude-plugin/plugin.json and a Claude marketplace at .claude-plugin/marketplace.json.
Inside Claude Code:
/plugin marketplace add 6Kmfi6HP/avoid-overkill
/plugin install avoid-overkill@avoid-overkill
/reload-plugins
For local development:
claude --plugin-dir ./avoid-overkillThis skill follows the Agent Skills directory shape:
skills/avoid-overkill/
SKILL.md
cases/
agents/
Copy skills/avoid-overkill into the skill directory supported by your agent client, or use the helper:
git clone https://github.com/6Kmfi6HP/avoid-overkill.git
cd avoid-overkill
./install.sh agents-userAsk your agent:
Use $avoid-overkill to review this code or paper and keep it proportional, evidence-grounded, and direct.
Tutorial 1 — Trim defensive checks in code
Use $avoid-overkill. Review this error handler: it retries 3 times, falls back to a cached value, logs 4 warning levels, and wraps every call site in a guard. Keep only what the contract requires.
The skill names the observed failure (none reproduced), removes unrequested retries/fallbacks/guards, keeps the one necessary boundary, and stops when user-visible behavior is correct.
Tutorial 2 — Stop test patching
Use $avoid-overkill. My test only passes after I added a mock for an internal helper. Is that right?
The skill flags the test-specific mock as implementation-detail coverage, points the test at public behavior, and only changes a test after showing its contract is wrong.
Tutorial 3 — Reuse before rebuild
Use $avoid-overkill. I wrote a custom retry-with-backoff class, but the project already depends on a library that does this. What should change?
The skill prefers the repository/dependency/standard-library implementation, removes the handwritten duplicate, and custom-builds only the unsupported part.
Tutorial 4 — Claim-centered paper prose
Use $avoid-overkill. Review my paper's abstract and conclusion: remove self-weakening phrases like "only", "limited", "unfortunately", and make the contribution explicit without overclaiming.
The skill opens with problem, gap, idea, hardest result; deletes evaluative negatives and imagined reviewer objections; states the supported advantage in prose; qualifies the exact boundary once.
教程 1 —— 去除代码里的防御性校验
使用 $avoid-overkill。审查这段错误处理代码：它重试 3 次、回退到缓存值、打了 4 层警告日志、每个调用点都套了防御判断。只保留契约真正要求的部分。
效果：点名实际出现的失败（没有就删），移除未要求的重试/回退/防御，只保留必要的边界，并在用户可见行为正确后停止。
教程 2 —— 停止测试打补丁
使用 $avoid-overkill。我的测试只有给内部 helper 加了一个 mock 才能通过，这样对吗？
效果：指出这是针对实现细节的补丁式覆盖，把测试引导到公共行为上；只有证明测试契约本身错误时才改测试，且测试与实现同步更新。
教程 3 —— 先复用再自建
使用 $avoid-overkill。我手写了一个带退避的重试类，但项目已有依赖库自带这个功能。应该怎么改？
效果：优先使用仓库/已有依赖/标准库/官方实现，删除手写重复代码，只自建真正缺失的部分。
教程 4 —— 以最强主张为中心写论文
使用 $avoid-overkill。审阅我的摘要和结论：删掉 "only"、"limited"、"unfortunately" 这类自我削弱表达，在不夸大的前提下把贡献写明确。
效果：以问题—缺口—想法—最难结果开篇，删除评价性否定和预写审稿意见，在正文中写明有证据支持的优势，边界只限定一次。
When the model shows any of these signs, remind it to use this skill:
- Adding
try/except, retries, fallbacks, or default values that are never triggered, to "defend" against imagined failures. - Adding parameter validation, type guards, or compatibility branches to every function "for robustness".
- Abstracting early, introducing interfaces or extra modules "for future expansion".
- Adding implementation-detail mocks or special cases to tests "to make them pass".
- Writing "our method is only preliminary", "still behind SOTA", or similar self-weakening lines in a paper.
MIT