ProjectsJuly 30, 2026

grill-loop

What it is

grill-loop is an explicit Agent Skill that routes the current task to the smallest useful capability, follows that capability's own contract, then reassesses from the result. It connects grilling, deep-grill, OpenSpec, visual taste, and other specialist skills without turning them into a fixed methodology.

Why it exists

Real work moves unevenly. A vague idea may need one focused conversation; a drafted plan may need autonomous adversarial review; a confirmed change may belong in OpenSpec; a frontend artifact may need a taste pass. A rigid pipeline either runs too much or starts owning artifacts that already have a better owner. grill-loop keeps only the routing decision. The selected capability keeps its own scope, authority, outputs, gates, and safety boundaries.

Contract

Choose one smallest useful move from the current goal and artifacts.
Carry only confirmed decisions, constraints, and artifact references into the handoff.
Skip, repeat, reorder, return, or stop as the work changes.
Add no shared schema, duplicate workflow state, hooks, background processes, or implicit invocation.

Install

Bash

npx skills@latest add nxxxsooo/grill-loop

Then say: Use $grill-loop to continue this task.

这是什么

grill-loop 是一个显式调用的 Agent Skill：观察当前目标与产物，只选择最有价值的下一项能力，遵循该能力自己的契约，再根据结果重新判断。它可以串联 grilling、deep-grill、OpenSpec、视觉 taste 与其他专家 Skill，但不会把它们改造成固定方法论。

为什么需要它

真实工作并不沿直线前进。模糊想法可能只需要一次聚焦访谈；成形方案需要自主对抗审查；确认后的长期变更适合进入 OpenSpec；前端产物则可能需要 taste 复核。固定流水线要么运行过多，要么开始接管已有明确所有者的产物。 grill-loop 只保留路由决定。被选择的能力继续拥有自己的范围、授权、输出、门禁和安全边界。

契约

从当前目标与产物出发，只选择一个最小且有价值的动作。
交接时只带上已确认决策、约束和产物引用。
随工作变化自由跳过、重复、重排、返回或停止。
不增加统一格式、重复状态、hooks、后台进程或隐式调用。

安装

Bash

npx skills@latest add nxxxsooo/grill-loop

然后说：使用 $grill-loop 继续这个任务。

Related projects

Codex Doubao Pets

六个可以直接安装到 Codex Desktop 的豆包桌宠。用真实角色图完成选择，用一条 npm 命令完成安装。

Read case study

GitHub

🚀 Landing Page

Slash Saver

一个只为物理斜杠键做一件事的原生 macOS 工具：在原始按键进入当前应用前，切到用户指定的 ASCII 输入源。

Read case study

GitHub

🚀 Landing Page

deep-grill

让 Agent 先拷问自己的 Agent Skill。自主调查方案每条分支、反驳自己的答案，只把真正主观的分歧一次性汇总给用户确认，确认前不动手实现。

Read case study

GitHub

Mingjian AI Services

A focused collaboration entry for teams moving from AI ideas to working products, Agent systems, and operable business workflows.

Read case study

🚀 Landing Page

opencode-sub2api-haiku-fix

A focused OpenCode plugin that fixes Claude Haiku extra-usage failures through Sub2API while preserving thinking and leaving Sub2API, OpenCode, and OMO untouched.

Read case study

GitHub

🚀 Landing Page