English

loop

一次触发，三个深循环，交付可验证结果。

loop 是唯一的常规入口。只需调用一次，它就会按产物就绪状态持续推进当前产品任务，不再要求你手动指定下一个 Skill。

想法 -> deep-grill -> Product Brief
     -> deep-design -> Build Contract
     -> deep-build -> 已验证结果

安装

推荐使用引导安装。它会安装完整 bundle，并让你选择 Agent、Project／Global 和 Symlink／Copy：

npx skills@latest add nxxxsooo/loop --skill '*'

需要直接完成全局安装时：

npx skills@latest add nxxxsooo/loop -g -y

快速安装会跳过提示，并按 skills CLI 的 Agent 检测结果全局安装完整 bundle。

开始任务时说：

使用 $loop 把这个想法推进到设计和实现。

后续普通回复会继续当前子循环。你也可以单独调用任意 deep-* Skill；直接调用不会自动开启持久 loop。

从 grill-loop v2 迁移

仓库已从 nxxxsooo/grill-loop 更名为 nxxxsooo/loop。GitHub 会重定向旧仓库 URL，但已安装快照仍保留旧 Skill 名。先删除一次，再安装 v3：

npx skills@latest remove grill-loop grilling -g -y

然后使用上面任一安装方式。

Skill	负责内容	就绪条件
loop	会话持续、产物迁移和 ?	用户要求的终点完成
deep-grill	想法澄清和对抗性审查	Product Brief 或审查结论确认
deep-design	产品规格与实现设计	Build Contract 确认
deep-build	实现、验证和授权范围内的交付	必过场景成功

? 表示解释并续接

在活跃 loop 中，只发送：

?

loop 会冻结当前动作，重新说明最后确认点、当前子循环和产物、发生的变化、重要性，以及原本准备执行的下一步，但不会结束当前 loop。

解释后，OMO 在可用时使用原生倒计时或自动续接；其他客户端优先显示本地化的原生问题界面，让你继续、调整下一步或停止。没有可调用的原生界面时，使用同语言的一句续接提示。选择继续会立即恢复原本下一步，不必重新点名任何 Skill。

原生问题工具

deep-grill 和 deep-design 会用当前客户端的原生问题界面处理归用户所有的决定。如果工具存在但受模式限制，子循环会保留待决问题，请你切换模式，然后等待。只有客户端没有原生问题界面，或者你明确选择文字方式时，才使用文字问题。

规格与设计

deep-design 负责完整的规格和设计过程。小型工作可以把 Build Contract 放在当前任务或原生计划中。耐久、多会话、跨组件或高后果变更，会使用项目的官方 OpenSpec 工作流作为实体 Build Contract。若存在该契约，deep-build 使用官方流程实施。

领域、架构、前端、测试和交付等专家能力由当前子循环按需调用，用户无需管理内部路由。

完整 loop 契约

以下内容与 skills/loop/SKILL.md 保持完全一致：

Loop

Own progression, not the child methods. Route the active product task through three artifact contracts:

deep-grill -> confirmed Product Brief
deep-design -> confirmed Build Contract
deep-build -> verified result

Keep The Loop Active

A fresh loop starts only when the user explicitly invokes loop or clearly asks to start this workflow. After that, treat every ordinary reply as a continuation of the same loop until the goal is achieved, the user stops, the user clearly changes tasks, or further work requires authority outside the original request. The user does not need to name loop or any child again.

Resume the current child while its output contract is incomplete. When its artifact is confirmed, select the next child from artifact readiness without asking the user to nominate a skill or choose a route. State the transition briefly and continue in the same response when the next action is already authorized. Artifact readiness chooses the child; it does not expand the user's authority or the requested endpoint.

Handle ? Without Breaking Flow

When the user's entire trimmed message is exactly ? during an active loop, freeze the current action at its exact frontier while keeping the loop active. Re-explain:

• the last confirmed point;
• the current child and artifact state;
• what changed;
• why it matters; and
• the next action that was about to occur.

Use the language of the user's latest substantive message and preserve established project terms. Do not delegate, advance an artifact, or present the child's pending domain decision while explaining.

Immediately after the explanation, present the best continuation control the active host actually supports:

• In OMO, use its native countdown or autoresume surface when available so the pending next action resumes automatically unless the user intervenes.
• Otherwise, when a native question interface is callable, ask whether to Continue (recommended), Adjust next action, or Stop; localize every label and explanation.
• When no native continuation interface is callable, ask one concise localized continuation question in prose. Recommend resuming the pending next action while accepting an adjustment or stop in free form.

Keep the current child, artifact, frontier, and authority active while the continuation control is open. Continue resumes the pending next action immediately. Adjust next action changes only that action unless the user's instruction invalidates an artifact. Stop ends the loop. Never require another skill invocation to resume.

Route By Artifact Readiness

• Use deep-grill when no confirmed Product Brief exists, when root product intent is unresolved, or when the requested endpoint is an isolated adversarial audit. Discovery mode produces the Product Brief; audit mode may finish without advancing when an audit is the whole request.
• Use deep-design when the Product Brief is confirmed but no confirmed Build Contract exists.
• Use deep-build when the Build Contract is confirmed and implementation is authorized.
• Resume the same child when its contract is incomplete, even after a mode switch, clarification, or ordinary user answer.
• Return an invalidated artifact to the child that produces it. A root product contradiction returns to deep-grill; a material behavior, interface, task, or verification gap returns to deep-design.
• Honor a direct invocation of a child skill. Direct use does not require loop, but it also does not activate the persistent loop unless the user explicitly starts one.

The active child selects any domain, architecture, research, interface, visual-design, testing, deployment, or other specialist it needs. The user does not manage internal routing.

Preserve One Source Of State

Carry the goal, current child, confirmed artifacts, constraints, authority, and evidence through the conversation and artifact references. For sustained work, use the active client's native goal or plan mechanism when available. Do not create duplicate workflow state, custom gates, hooks, or background processes.

Never imitate a child contract inside loop. Delegate the work, accept its artifact only when its readiness rules are met, and reroute from evidence. Stop when the requested endpoint is complete, the user stops or changes tasks, progress requires new authority, or remaining work has diminishing returns.

Raycast

导入 skills/loop/assets/raycast-snippets.json，即可使用 ;lp、;dg、;dd 和 ;db。安装 Skill 不会修改 Raycast。

更新

已安装副本是快照：

npx skills@latest update loop deep-grill deep-design deep-build -g -y

如果客户端已经缓存 Skill 元数据，请重新加载或新建任务。

许可证

本 bundle 使用 MIT License。deep-grill 保留已整合上游源码的归属信息，见 THIRD_PARTY_NOTICES.md。