菜单
Collaboration protocol for Ralph loop (Plan → Act → Reflect → Refine). Use when working from spec files in specs/<name>/ (goal.md/spec.md, plan.md, state.json, decisions.md); when executing tasks in a shared plan; or when the user mentions Ralph, multi-agent, or file-based collaboration.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
基于 SOC 职业分类
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/GetStream/stream-chat-react --skill ralph-protocol命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Maintain WCAG-focused accessibility in stream-chat-react. Use when changing interactive components, dialogs, menus, forms, media controls, notifications, focus behavior, keyboard flows, aria attributes, or screen-reader announcements.
Styling and structure conventions for stream-chat-react. Use when adding or editing components, SCSS, or icons in this repo—file layout, styling folder structure, SCSS imports, and icon placement.
Structures plan.md files for parallel, non-overlapping agent work. Use when creating or editing execution plans, task lists for multiple agents, or when the user asks how to structure plans or split work into tasks.
| name | ralph-protocol |
| description | Collaboration protocol for Ralph loop (Plan → Act → Reflect → Refine). Use when working from spec files in specs/<name>/ (goal.md/spec.md, plan.md, state.json, decisions.md); when executing tasks in a shared plan; or when the user mentions Ralph, multi-agent, or file-based collaboration. |
Files are the source of truth. All agents share memory via files. No silent decisions. Always update state after acting. Prefer small, explicit updates.
Location: Generate and keep spec files in a dedicated folder inside the specs folder at the root of the repo: specs/<feature-or-component-name>/. Example: specs/resizable-container/goal.md, specs/resizable-container/plan.md, specs/resizable-container/state.json, specs/resizable-container/decisions.md. One folder per feature or component; do not put spec files next to source (e.g. not under src/components/...).
| File | Purpose |
|---|---|
| spec.md or goal.md | What we achieve; success criteria; constraints; non-goals. Read first. Only change if goal actually changes. No implementation details. |
| plan.md | How we achieve it. Ordered tasks, ownership, dependencies, status (pending | in-progress | done | blocked). Propose changes before big deviations; don't rewrite completed sections. |
| state.json | Current memory. Task statuses, flags (blocked, needs-review, etc.). Update immediately after acting. Read before assuming anything. |
| decisions.md | Log of non-trivial decisions (what + why). Append only; never delete. Prevents reopening or contradicting past choices. |
Before acting: Read the spec files in specs/<name>/: spec.md (or goal.md) → plan.md → state.json → decisions.md.
During: Follow the plan; no overlapping work unless coordinated; no undocumented decisions.
After: Update specs/<name>/state.json → record decisions in specs/<name>/decisions.md → update task status in specs/<name>/plan.md. Optionally add learnings to observations.md in the same folder.
Prohibited: Decisions without recording; using chat as memory; re-solving done problems; changing goals implicitly; overwriting files without explanation.
specs/<name>/plan.md), marked in-progress and assigned to you.Only commit files that were changed for the current task. Use git add <file> — never git add -A or git add .. Leave unrelated changes unstaged.
When acceptance criteria involve UI: use Playwright (MCP or project config). Take screenshots for success and failure. Save to a clear location (e.g. __tests__/screenshots/).
Each iteration: Plan (update specs/<name>/plan.md if needed) → Act (scoped work) → Reflect (learnings) → Refine (plan or decisions).
For decision log format and state.json example, see reference.md. Plan structure and worktrees: use make-plans and worktrees skills.