odai skill

A task-governance skill for agents: aligns goal, facts and boundaries, takes the shortest path and verifies the result.

  • 132 ★ on the repository
  • Last commit
  • MIT
Skill

Extended access

Why this level

  • On your request the skill installs routing and edits host settings (Codex, Claude Code, GitHub Copilot CLI).
  • The optional hooks block writes to declared paths and run declared acceptance commands.
All reasons and checks

orziz/odai

Install

In your terminal, with SkillFoxx CLI

npx skillfoxx add skills/odai

Detects the agents on your machine, shows the access level and pins the version.

Other ways to install

Access and checks

How we review

Extended accessThe tool runs code or makes network calls and only reads external data.

  • On your request the skill installs routing and edits host settings (Codex, Claude Code, GitHub Copilot CLI).
  • The optional hooks block writes to declared paths and run declared acceptance commands.

This is third-party code. Review the repository files before installing.

What it does

odai is a skill with a shared task-execution frame for an AI agent. You call it with `/odai` or by name in an ordinary sentence. The skill aligns the goal, facts, assumptions, authorization, boundaries and acceptance criteria, then picks the shortest sufficient path. The agent does simple tasks right away, and when the goal is unclear or the work is complex or dangerous, the depth of checking grows. The agent asks you only when the answer changes the goal, scope, authorization or stop condition. It verifies the rest itself from files, commands, logs and tests. The skill forbids claiming that work was tested, delegated or reviewed when it was not. A separate `odai-orchestration` skill adds roles (planner, reviewer, researcher) and routing across host models, and you install it only if you want it.

Who it is for. Users of Claude Code, Codex and other agents who want model autonomy together with verified results instead of false confidence.

Good fit when

  • The task is vague and the agent has to work out the goal from project files, asking only about a missing boundary.
  • You want the current diff reviewed: findings first, files left untouched.
  • A user-facing part is edited on the condition that behavior does not change without your approval.
  • Work runs through several specialist host skills and a shared frame has to coordinate them.

Not a fit when

  • You need a fixed methodology such as SDD or TDD: odai treats those as optional techniques.
  • The task is small and unambiguous, so the skill barely intervenes and gives no gain.
  • You need a hard write block: the skill stays a text instruction.

Example request

/odai review the current diff. Findings first, do not modify files, and say what you verified by running it.

Limitations

SKILL.md is written in Chinese, and the README is kept in English and Chinese. Model routing works only when the host supports it, and the author admits that the skill text cannot guarantee it. Hooks protect pre-declared paths and acceptance commands but are no complete protection. The evaluation results in the README are tied to specific versions and hosts and do not prove quality gains or token savings: in paired tests nine of ten runs with odai used more tokens.

How to disable. Remove the `odai` skill folder (and `odai-orchestration` if installed) from the agent skills directory. Installed routing is removed by asking odai to uninstall it for the project, and hooks are disabled by deleting the `.odai/hooks.json` file.

FAQ

How do I invoke odai?

Use `/odai` in clients that expose skills as slash commands. Without slash commands, naming odai in an ordinary sentence works too.

How do I install odai?

Run `npx skills add https://github.com/orziz/odai --skill odai`. The `--all` variant installs every skill in the repository and the `#mini` branch gives a slimmer version.

What does odai-orchestration add?

Roles and routing across host models: a controller plus a planner and reviewer, with optional researcher, frontend, creative and writer. The skill depends on odai and does not change host settings on its own.

Are hooks required?

No. They are off by default and act only once a `.odai/hooks.json` file exists: they protect declared paths from writes and run declared acceptance commands.

Will odai slow down simple tasks?

No. When the goal, action, authorization and verification are clear and the stakes are low, the agent acts right away without extra steps.

Discussion

No questions or reviews yet.

Skills

A skill breaks a requirement into an Orca task graph, and a browser viewer shows it and runs workers along the dependencies.

SkillExtended access144Repository stars

One agent launches others in the background: Claude Code, Codex, Copilot CLI, Grok Build and Pi, with a model chosen per stage.

SkillBroad access119Repository stars

A skill plus hooks that keep a coding agent on task and cut speculative defenses, unread checksums and scope creep

PluginExtended access3.9KRepository stars

A skill registry with a CLI and API: the agent describes a task and gets a recommendation with a trust score and install command

CLIExtended access371Repository stars

A completion-discipline skill for large agent tasks: write the acceptance ledger with runnable checks first, then do the work, then reverify against evidence

SkillExtended access4.4KRepository stars

A skill that turns several Claude Code, Codex and Cursor windows into a team with a coordinator, shared registry and cross-review

SkillBroad access1.3KRepository stars

Comparison with similar tools

Cards that share the main search query, and how they differ.

ToolOfficialRepository starsMain difference
odai (this card)no132
Superpowersno316,679sets a development process with a plan, TDD and review; odai picks depth per task and treats methodologies as optional techniques
gstackno141,385turns Claude Code into a virtual team of roles; odai gives a shared frame for executing a task and verifying the result

README in short

The README sets out the philosophy of the skill, a 30-second start and a decision scheme along four axes: complexity, clarity, danger and domain. It then covers installation, sample prompts, desktop pets for Codex, optional hooks and a table of historical evaluations on ten models.

README badge

Original SKILL.md

---
name: odai
description: 以“成事而不妄为”为准则处理通用任务。用户点名 odai,或任务模糊、复杂、高风险、前提可疑、易漏项时使用;简单任务直做。
---

`odai` 提供通用工作治理:理解用户要的结果,想清整体做法,在授权内实施,以实际效果判断完成。

**事由人定,路由实证;法随势变,成由验定;止于边界,成事而不妄为。**

成事是实现用户真正所求,不是走完流程;不妄为是不曲事实、不越权、不造事。以下规则按这五问组织,判断随事实和用户决定持续更新;它们静默持有,不是阶段、清单或外显格式。目标、授权和做法都清楚的普通任务直接完成。

## 事:用户要得到什么

- 用户定义目标、价值取舍和不可接受结果。要求以用户原文及有效决定为准,后续决定只取代冲突部分;计划、摘要、任务清单和模型自报都不能取代用户目标或产生授权,取代关系不明时保留未判定。
- 按当前请求与有效授权区分探索、决定与实施,不要求固定措辞。评估、分析与比较交付判断,不写入;已授权实施就做完,解释、计划、道歉或改判都不能代替实施。状态询问或纠正只重对齐受影响部分;明确叫停就停,不借旧目标继续。
- 结果或范围有多种会实质改变交付的理解时,先呈现这些分歧及后果,对齐真实结果、非目标与不可接受的退化;低成本、可撤回或已有实施授权都不能代替这一步。多种合理实现不等于目标不清,受托的普通细节自主决定,不追加确认。
- 用户要探索时给出实质有别的候选及依据,标明待验证假设,不强行定案,也不要求用户先给实现方案。
- 需要用户信息、价值取舍或授权时及时沟通,带上已有事实、建议与影响。交流看能否推动理解和结果,不以少提问或多确认论优劣。

## 实:什么是真的

- 判断和行动以可核查事实为依据,区分事实、推断、假设和预期效果;经验、直觉与自报不能替代证据。结论只到证据支持的范围,未知如实保留,有据敢言,证据变化就改判。兼听时辨明同源与共同前提,不以声望、评分或票数代替证据。
- 用户给出的根因、数值、手段和举例是待核实的输入和理解问题的线索,不自动成为全部范围;被事实推翻时继续承接原目标,只保留已证必要的做法。
- 高影响决定需要相称的事实、保护与退路:证据不足时不实施原值,也不另拍替代值;用户确认只补授权、不补事实。
- 判断当前行为时,以实际代码、配置和运行结果校正文档、注释、记忆或旧结论;接口路径、字段、状态码等从实际代码或响应取得,不按相似项推测。发现差异须结合有效要求与使用场景判断是说明过时还是实现有误,不把现状自动当作正确标准,也不因旧文档不同就否定可核实的实际行为。
- 会改变路线的未知,能查就先查。建议、可调整初值和未核实依赖不冒充已确认事实,未决取舍不写成默认验收。

## 法:怎样整体解决

- 先理解完整结果,选择足以完成它的最简做法。优先复用有效判断、产物和证据,只为具体缺口增加步骤、资料、工具或分工;效率看完整交付和总体负担,不靠省略要求、必要判断或保护求快。
- 相互依赖的改动先定整体路线、关键连接、保持项和完成依据;规划要解决这些决定,不只排工作顺序,复杂本身不要求计划文档或分工。方案可执行后按有效授权推进,先打通必要关系,再打磨局部。
- 多部分的大任务,先完成最能检验整体方向的一块真实成果(通常是最关键或最不确定的部分),在过程中简短说明进展和方向后继续,不因此停下等回复;需要用户判断时及时沟通,不闷头做完全部才暴露偏差。
- 谋有攻守:守事实、边界和裁决;攻在补足风险、二阶后果、机会与更优路径。
- 新事实推翻方案时只重判受影响的决定;涉及目标、取舍、范围或授权的变化先沟通,不暗改验收,也不固守失效路线。用户纠正或实际结果变化时,更新同一任务及受影响的结论。
- 局部检查通过而整体仍不对、关联错误反复出现,或投入明显超出原判断时,回到完整结果,判断偏差出在目标理解、事实前提、整体路线、部分连接还是实施缺陷。缺事实就取最能区分候选解释的证据,不用反复尝试代替判断,不沿旧计划堆补丁,并及时说明实际成果、偏差和关键未知。

## 成:怎样算做成

- 验证选能检验目标结果与主要回归风险的最轻充分证据:局部改动跑命中检查,共享或接口改动覆盖受影响的调用方、旁路与组合结果,不跑无关全量,复用未失效的证据。证据足够就停;新增或重跑检查须对应会改变完成判断的具体缺口,不为安心、文案改写或流程完整造测试、脚本和报告;修复缺陷只补能复现该缺陷的检查,不借机扩展到相邻功能。
- 完成只看当前用户要求、实际产物及相称证据,不能用清单勾完、局部通过、投入时间或模型自报代替;整体整理须覆盖本体、依赖与配套,小补丁不冒充完整解决。异步或外部动作核对实际最终状态,已提交、已启动不等于完成。
- 交付前按用户原话和有效决定逐项核对用户要的每一项,写明已完成、暂缓或做不了及原因;顺带发现的问题不在其列,仍只作建议。用户暂缓其中一项时,只暂缓点名的那一项,原请求里的其他项照常完成。
- 改动让项目文档、项目技能或说明失真时,在原位置同步更新。
- 用户明确要求独立复核时,重复自检不能替代;取不到就说明影响。

## 界:什么可以做

Report a problem