Skip to content
Back to search

private-house-code-v2.5

让 Codex 少烧 Token、少造复杂度,同时不降智的代码约束 Skill

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/see-sol-lab--private-house-code-v2-5/badge.svg)](/listings/see-sol-lab--private-house-code-v2-5)

HTML

<a href="/listings/see-sol-lab--private-house-code-v2-5"><img src="/api/listings/see-sol-lab--private-house-code-v2-5/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 95 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: private-house-code
  3. description: Use for every task that plans, writes, edits, debugs, tests, refactors, reviews, or maintains code. Apply to professional repositories, scripts, apps, services, prototypes, beginner work, vibe coding, and bounded code changes inside large systems. Choose the fewest clear lines, files, paths, states, and abstractions that correctly implement the requested behavior. Do not invent fallbacks, compatibility paths, defensive layers, or future scale without evidence. Do not use for ordinary conversation, emotional support, prose, translation, summarization, social-media content, or general research or technical discussion that is not part of planning, changing, or reviewing code. Never weaken real security, integrity, concurrency, compatibility, privacy, accessibility, legal, or regulatory requirements.
  4. ---
  5. # Private House Code
  6. Write the feature. Do not build a bank around it.
  7. For coding tasks, build the **smallest complete implementation**: correct at its real boundaries, obvious to a human maintainer, aligned with the current project, and no more complex than the requested behavior requires.
  8. **Every added line, file, abstraction, state, path, dependency, process, check, and fallback must have a present job.**
  9. ## 1. Apply only to coding, then take the direct path
  10. If the current task does not plan, change, test, review, or maintain code, do not apply the remaining instructions.
  11. For coding work:
  12. - Start from the requested behavior, not an imagined architecture.
  13. - If one readable line is enough, write one readable line. If one ordinary function is enough, do not create a class, interface, service, manager, adapter, or factory.
  14. - When several designs are equally correct, choose the one with fewer moving parts overall, not merely fewer lines. A shorter implementation is not simpler if it introduces a heavier persistence model, resource lifecycle, migration burden, test scaffolding, or failure modes the current task does not need.
  15. - For a new single-user local tool, start with the simplest persistence that fully supports the requested behavior. Do not escalate storage merely for generic robustness, future flexibility, or abstract data-integrity advantages; use a heavier store only when real concurrency, transactions, query needs, data scale, or an existing project contract requires it.
  16. - Prefer one production path and one source of truth. Derive secondary values instead of storing duplicates.
  17. - Reuse the standard library, current modules, existing dependencies, and project conventions before adding anything new.
  18. - Do not code-golf. Shorter is better only while the code remains normal, readable, debuggable, and easy to change.
  19. - Do not add features, extensibility, cleanup, or redesign that the user did not request.
  20. - In a large repository, respect existing contracts and required checks while keeping a bounded change bounded.
  21. ## 2. Do not invent enemies or fallbacks
  22. The task begins as “implement the requested behavior,” not “search for hypothetical attackers, outages, formats, users, or future scale.”
  23. Without an explicit requirement or present evidence, do not add:
  24. - fallback chains, secondary truth sources, alternate paths, dual reads, dual writes, dual runtimes, or multiple parsers
  25. - caches, snapshots, replicas, legacy recovery, blanket retries, backoff, circuit breakers, failover, or self-healing loops
  26. - queues, polling, schedulers, background workers, watchdogs, speculative compatibility, extension points, distributed coordination, or lifecycle-state machinery
  27. “Safer,” “best practice,” “enterprise-ready,” and “we may need it later” are not evidence by themselves.
  28. A second path is justified only by a current user requirement, supported compatibility contract, observed failure, existing system dependency, or concrete credible risk. When justified, define its caller, precedence, failure behavior, and—if temporary—removal condition.
  29. Prefer a direct, useful error over machinery that guesses, silently switches sources, or pretends success.
  30. ## 3. Preserve protection for real boundaries
  31. Keep the smallest correct control for:
  32. - authentication, authorization, secrets, and public or untrusted input
  33. - path boundaries, destructive actions, and irreversible deletion
  34. - atomic writes, transactions, data-integrity controls, and concurrent-access mechanisms when the current operation actually requires them
  35. - cleanup, cancellation, and timeouts for actual external operations
  36. - released API, file-format, migration, and version compatibility
  37. - privacy, accessibility, audit, safety, legal, and regulatory obligations
  38. Security follows the real threat model, not repository size. Protect the boundary that exists; do not turn unrelated code into a bank.
  39. Follow established shared infrastructure when the current path depends on it. Do not create a parallel path, and do not remove paid-for infrastructure merely to satisfy the metaphor.
  40. ## 4. Spend less code, work, and token
  41. - Keep calibration internal. Do not announce this Skill, recite private-house/workshop/bank categories, repeat the request, or produce a risk matrix for an ordinary task.
  42. - When the task is clear, act directly. Ask one focused question only when the missing answer would materially change behavior, safety, or architecture.
  43. - Read, search, edit, and test only what can affect the result. Do not use tools merely to appear thorough.
  44. - Run the smallest relevant check first; expand only when project requirements or evidence justify it.
  45. - Do not hash files as a routine per-turn completion ritual. Compute or compare file/content hashes only when byte identity or integrity is the thing being verified, when syncing exact external artifacts, or at a meaningful task/release checkpoint where hashing adds evidence. During an ongoing bounded task, prefer focused tests, diffs, or status checks and do not repeat the same hash verification after each conversational turn.
  46. - Do not write a long plan unless the user asks or the task is genuinely architectural.
  47. - Keep the completion report brief: what changed, what was checked, and any unusual fallback, compatibility path, retry, background process, or extra state that remains and why.
  48. ## 5. Leave code a human can maintain
  49. - Keep the input → processing → output path easy to trace.
  50. - Use clear names, direct control flow, and honest errors.
  51. - Avoid clever indirection, pass-through layers, hidden second sources, and defensive branches with no demonstrated purpose.
  52. - Preserve user-visible behavior unless the task changes it.
  53. - Touch only required files and concepts; do not hide broad cleanup or redesign inside a focused change.
  54. - Remove replaced code and orphans when no active compatibility obligation needs them.
  55. - Test user-visible behavior, real boundaries, and data integrity—not an imaginary internal protocol.
  56. - Never make a failing check green by skipping, weakening, or mocking away the failure.
  57. Before finishing, ask of every addition:
  58. 1. Which present requirement or current caller uses it?
  59. 2. Which observed failure, existing contract, or credible risk requires it?
  60. 3. Why is a simpler correct implementation insufficient?
  61. 4. Can a human maintainer understand it without unnecessary indirection?
  62. If it cannot answer, remove, merge, or postpone it.
  63. Read `English Test Records and Test Cases/references/calibration-examples.md` only when a boundary is genuinely ambiguous.
  64. ---
  65. > **Project provenance / 项目来源:** Private House Code was originally created and is maintained by [See-Sol-Lab](https://github.com/See-Sol-Lab). Canonical repository / 官方原始仓库: [private-house-code-v2.5](https://github.com/See-Sol-Lab/private-house-code-v2.5). Licensed under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).
README.md · README · 256 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. <p align="center">
  2. <img src="assets/private-house-code-hero.png" alt="Private House Code — Write the feature. Do not build a bank around it." width="100%">
  3. </p>
  4. <h1 align="center">Private House Code v2.5|私宅级代码约束</h1>
  5. <p align="center">
  6. <strong>Write the feature. Do not build a bank around it.</strong><br>
  7. <strong>写一个功能就写功能,不要在它外面修一座银行金库。</strong>
  8. </p>
  9. <p align="center">
  10. <img src="https://img.shields.io/badge/release-v2.5-5965F3?style=flat-square" alt="Release v2.5">
  11. <a href="LICENSE"><img src="https://img.shields.io/badge/license-CC_BY--NC--SA_4.0-F57C00?style=flat-square" alt="CC BY-NC-SA 4.0"></a>
  12. <img src="https://img.shields.io/badge/Codex-Skill-111827?style=flat-square&logo=openai&logoColor=white" alt="Codex Skill">
  13. <img src="https://img.shields.io/badge/evaluation-reproducible-0F766E?style=flat-square" alt="Reproducible evaluation">
  14. <a href="https://doi.org/10.5281/zenodo.21836206"><img src="https://zenodo.org/badge/DOI/10.5281/zenodo.21836206.svg" alt="DOI 10.5281/zenodo.21836206"></a>
  15. </p>
  16. <p align="center">
  17. <a href="SKILL.md">v2.5 SKILL.md</a> ·
  18. <a href="gpt-6-astra/">GPT-6 Astra 适配版</a> ·
  19. <a href="测试记录(供复现)/">中文测试记录</a> ·
  20. <a href="English%20Test%20Records%20and%20Test%20Cases/">English evaluation</a> ·
  21. <a href="与Ponytail的查重报告.md">Ponytail 查重报告</a> ·
  22. <a href="https://doi.org/10.5281/zenodo.21836206">Zenodo archive</a>
  23. </p>
  24. ## 2026-09-07 更新:GPT-6 Astra 适配版
  25. 新增 [`gpt-6-astra/SKILL.md`](gpt-6-astra/SKILL.md):更短、更直接地表达私宅级要求,减少冗余设计、测试、哈希和审批步骤,同时完整完成任务并保留真实边界。
  26. **本次为提示词适配更新,尚未进行 Astra 行为测试或对照评测。** 下文 v2.5 的测试结果和费用数据仍属于 GPT-5.6 Sol High,不能外推到 Astra。两版二选一安装,详见[适配说明](gpt-6-astra/README.md)。
  27. Added a shorter [GPT-6 Astra adaptation](gpt-6-astra/). It has not undergone Astra behavioral testing or paired evaluation; existing v2.5 results apply to GPT-5.6 Sol High. Install one variant at a time. The sections below describe the original v2.5 release.
  28. ## v2.5更新
  29. Private House Code 是一个适合**全局加载**的日常编码 Skill,根据实测,装skill后的费用总消耗比原装codex低约 47.9%(数据见报告)。
  30. **写这玩意最初时间是7月20号,我实在是受不了5.6烧额度了,1步10哈希,一轮50测试,真是受不了那个fallback**
  31. 所以写完自用了半个月,效果很好,遂公开。
  32. ## 创作理念:
  33. 1. 写一个小功能,就不要整成银行金库
  34. 2. 能一行实现的就一行
  35. 3. 但是一行实现不了的,不许刻意砍
  36. 4. 执行skill,且不降质、不降智
  37. 5. git上面高星的类似项目有没有已经写好的?8.6 搜了一次,暂时没有
  38. 6. 多快好省 + 高质量完成
  39. 7. 傻瓜式,大佬和小白都能用,至于用了怎么改我就不管了,大佬建议自改来适应全局,小白建议别自己乱改
  40. 8. 企业级巨型架构不适用,写大项目的别用啊,除非你改的很好+测试没问题再用
  41. 项目里面有一整套测试题,包括测试情况,测试费用,题目,记录等等,感兴趣的可以自己去跑或者复现啊,
  42. 总之节约近一半是真的。
  43. 为什么版本号是V2.5呢,因为7.20自用的版本是V1.0,8.4优化了第二版,XHS有网友希望我分享,出于负责和严谨,昨天晚上出了那套测试题先跑V1.0和V2.0的smoke,发现两句有点小问题,重新调整后变成现在的V2.5;
  44. 8.7早上开始又再重跑一次全套测试,确认得到的数据没问题、没有影响5.6,而且效果非常好,于是封装+公开仓库了。
  45. 该Skill适用于:
  46. - 专业程序员的日常功能、修复、维护和局部重构
  47. - 初学者与非程序员的 vibe coding
  48. - 个人应用、脚本、CLI、本地工具和原型
  49. - 内部工具、小团队服务
  50. - 大型代码库中边界明确的小改动
  51. 以下情况不适用:
  52. - 真正的企业级架构设计
  53. - 公共多租户系统
  54. - 金钱、健康、安全、法律与不可逆损失
  55. - 身份认证、权限、密钥和账户恢复等真实安全边界
  56. - 其他巨型架构和公有项目
  57. ## Related Work / 相关项目
  58. 经社区成员提醒,我们对 [Ponytail](https://github.com/DietrichGebert/ponytail) 这个“约束 AI 过度工程”问题域中的重要先行项目,和我们的Skill相似度、差异进行深入复核。两者共享 YAGNI、优先复用现有能力、避免无需求抽象并保留真实安全边界等工程原则,但目标和机制不同:
  59. Ponytail 以“需求必要性 → 复用 → 标准库 → 平台原生能力 → 已有依赖 → 最小实现”的方案选择梯子为核心;
  60. Private House Code 则以“每一层复杂度必须由当前需求、现有契约、已观察故障或可信风险付费”为核心,重点约束 GPT-5.6 Sol/Codex 的 fallback、双路径、重复事实源、状态机、过度搜索/测试/哈希和多轮停止行为。Private House Code 为独立开发项目,未复用 Ponytail 的文本或规则结构。
  61. 两个项目的目标和思路并不完全一致,欢迎对项目重叠度进行公开审阅。
  62. 详细对比见根目录的查重报告。
  63. ## English Description
  64. **Private House Code is a global coding Skill for keeping AI-generated code complete, maintainable, and proportionate to the real task.**
  65. Private House Code is a global coding Skill for Codex. It keeps ordinary feature work direct, maintainable, and proportionate to the task that actually exists.
  66. Agentic coding has a common failure mode: uncertainty turns into architecture. One feature becomes another abstraction, another source of truth, another fallback, another state, another worker, and another system a human must maintain.
  67. Private House Code changes the default:
  68. > **Every piece of complexity must be paid for by a current requirement, an observed failure, an existing contract, or a credible risk.**
  69. **Status:** Public release v2.5 · DOI: [10.5281/zenodo.21836206](https://doi.org/10.5281/zenodo.21836206).
  70. ## The five product promises
  71. ### 1. Coding only, but all coding
  72. Route it for planning, writing, editing, debugging, testing, refactoring, reviewing, and maintaining code—from professional repositories to scripts, prototypes, beginner work, vibe coding, and bounded changes inside large systems.
  73. Do not route it for ordinary conversation, emotional support, prose, translation, summarization, social-media writing, or general research and technical discussion that does not plan, change, or review code. Its engineering style must not bleed into normal dialogue.
  74. ### 2. The fewest clear lines that correctly solve the task
  75. If one readable line is enough, use one readable line. If one function is enough, do not build a class hierarchy. If one path is enough, do not create several.
  76. This is not code golf. The target is the least conceptual and maintenance weight while preserving normal readability, debugging, correctness, and project conventions.
  77. ### 3. No unpaid fallback machinery
  78. Push back on fallback chains, dual reads and writes, multiple parsers, legacy recovery, blanket retries, caches, recovery workers, and defensive state machines when no present requirement or evidence pays for them.
  79. A direct implementation that works does not need several imaginary escape routes around it.
  80. ### 4. Less code and less token waste
  81. Keep routine calibration internal. Clear tasks should be executed directly, without an airport-security checklist, repeated requirements, speculative risk matrices, oversized plans, or unrelated tool calls.
  82. Use the smallest relevant reads, edits, and tests first. Expand only when the repository or evidence requires it.
  83. ### 5. Code a human can understand and maintain
  84. Keep the main path easy to trace from input to processing to output. Use clear names, honest errors, explicit data sources, and abstractions with a current job.
  85. The task begins as “implement the requested behavior,” not “search for imaginary enemies.”
  86. ## Who this is for
  87. - professional developers using Codex for features, fixes, maintenance, review, and refactoring
  88. - beginners who need working code without an accidental enterprise-architecture course
  89. - non-programmers and low-code users doing vibe coding
  90. - scripts, CLIs, desktop tools, personal apps, prototypes, and local automation
  91. - internal utilities and small-team services
  92. - bounded changes inside large professional repositories
  93. A repository does not need to be small. A small change inside a large system is still a small change.
  94. ## Global default, evidence-based exception
  95. The default assumption is:
  96. > **This is a feature, repair, or bounded change—not a bank-scale architecture project.**
  97. The Skill still preserves real requirements at real boundaries: authentication, secrets, public input, path safety, irreversible operations, transactions, actual concurrency, released compatibility, money, health, safety, privacy, accessibility, legal, audit, regulatory, distributed, and availability obligations that genuinely apply.
  98. A trusted local settings file does not need a fortress. A public password-reset endpoint does not become safe because the app is small.
  99. Even inside a true bank-scale system, the Skill can keep a local change surgical. It protects the vault without making every drawer a blast door.
  100. ## What it blocks by default
  101. Unless the current task genuinely needs them:
  102. - fallback chains, secondary truth sources, dual reads, dual writes, and dual runtimes
  103. - generic repository/service/manager/adapter/factory stacks for one implementation
  104. - unused plugin systems, strategy registries, and future-proof extension points
  105. - queues, polling, schedulers, background workers, and recovery services without a current job
  106. - caches, snapshots, blanket retries, failover, and self-healing state machines without evidence
  107. - distributed coordination in non-distributed work
  108. - lifecycle-state explosions for failures the product does not handle
  109. - drive-by refactors hidden inside a focused feature or bug fix
  110. ## Routing and the 100% limit
  111. `SKILL.md` uses broad positive coding triggers plus explicit non-code exclusions. That is the strongest routing signal the package itself can provide.
  112. Skill text alone cannot guarantee selection on every turn; final invocation belongs to the Codex host and its Skill router. This repository therefore does **not** claim a technical 100% trigger guarantee.
  113. For best routing:
  114. - install it as a global Skill available to coding sessions
  115. - keep the Skill name and frontmatter description intact
  116. - avoid duplicate or conflicting versions
  117. - do not replace it with a global personality instruction that would pollute non-code conversation
  118. The V1.0 baseline was used globally in sustained local development. The public v2.5 release was tested separately against that frozen baseline and includes reproducible evaluation records in this repository.
  119. ## Examples
  120. ### One-line utility
  121. Add `is_blank(value: str) -> bool` to an existing Python utility module.
  122. Expected: one ordinary function with one readable return expression and only the focused test normal for the project—not a validator class, strategy interface, regex subsystem, or configuration layer.
  123. ### Local export
  124. Add CSV export to a one-user desktop app.
  125. Expected: reuse existing data, use the standard CSV library, choose one destination path, report write failure clearly, and add the smallest repeatable check—not an export framework, queue, fallback store, plugin registry, or lifecycle state machine.
  126. ### Public password reset
  127. Expected: secure random tokens, expiry, one-time use, authorization, abuse controls, and tests. Wrong use: removing real security because the app or user count is small.
  128. More boundary cases live in [`calibration-examples.md`](English%20Test%20Records%20and%20Test%20Cases/references/calibration-examples.md).
  129. ## Package and installation
  130. ```text
  131. private-house-code-v2.5/
  132. ├── SKILL.md
  133. ├── README.md
  134. ├── gpt-6-astra/
  135. │ ├── SKILL.md
  136. │ └── README.md
  137. ├── CITATION.cff
  138. ├── LICENSE
  139. ├── 安装skill的全局配套.md
  140. ├── English Test Records and Test Cases/
  141. │ ├── references/
  142. │ │ └── calibration-examples.md
  143. │ └── evals/
  144. │ ├── pressure-cases.md
  145. │ └── v2.5-sol-5.6-high-report-2026-08-07.md
  146. └── 测试记录(供复现)/
  147. ├── V2.5测试报告-中文版.md
  148. ├── 1-5题带SKILL测试记录.md
  149. ├── 1-5题裸机(卸载skill).md
  150. ├── 10-14题带SKILL测试记录.md
  151. ├── 10-14题裸机(卸载skill).md
  152. ├── 中间几题.txt
  153. ├── 1.png
  154. ├── 2.png
  155. ├── 3.png
  156. └── 4.png
  157. ```
  158. For v2.5, install the root `SKILL.md` together with `English Test Records and Test Cases/references/`. For GPT-6 Astra, install only `gpt-6-astra/SKILL.md`. Use the `private-house-code` skill directory in the location supported by your current Codex surface, then reload that surface. Choose one variant; avoid copying this entire repository into a skill discovery directory, which may expose both versions. See [`安装skill的全局配套.md`](安装skill的全局配套.md) for the companion global-install notes.
  159. Codex installation paths and host behavior can vary by surface and build, so this README does not claim one universal filesystem path.
  160. ## Evaluation and portability
  161. [`pressure-cases.md`](English%20Test%20Records%20and%20Test%20Cases/evals/pressure-cases.md) covers non-code exclusion, coding routing, readable one-line implementation, vibe coding, concurrency, authentication, destructive operations, external APIs, distributed platforms, state-machine bloat, compatibility, large repositories, and interaction efficiency.
  162. The v2.5 release is designed to reduce speculative complexity and token overhead without weakening real safety.
  163. The package is authored as a Codex-style Agent Skill. Its principles are language- and model-agnostic and may be adapted for Claude Code, GitHub Copilot, `AGENTS.md`, or other reusable-instruction systems. The v2.5 release ships one tested Codex package rather than several speculative wrappers.
  164. ## Version lineage
  165. V1.0 is frozen in a separate private archive as the original production-used baseline. V2.0 was an intermediate test candidate. After smoke testing identified two wording issues, the Skill was adjusted and re-tested as v2.5. This repository contains the public v2.5 release; the V1.0 baseline remains separate to avoid accidental duplicate loading.
  166. ## License and attribution / 许可与署名
  167. Copyright © 2026 [See-Sol-Lab](https://github.com/See-Sol-Lab).
  168. This repository is licensed under the [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License](LICENSE). You may use, share, and adapt the material with attribution for non-commercial purposes. Adapted material must use the same license, and changes must be indicated.
  169. 本仓库采用[知识共享署名—非商业性使用—相同方式共享 4.0 国际许可协议](LICENSE)。允许在保留署名的前提下非商业使用、分享与改编;改编版本须采用相同许可,并注明所作修改。
  170. ---

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.