一键导入
meta-sk-agent-writer
Use when creating or editing agent.md files for Claude Code or Pi subagents. Not for SKILL.md — use meta-sk-skill-writer instead.
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。
菜单
Use when creating or editing agent.md files for Claude Code or Pi subagents. Not for SKILL.md — use meta-sk-skill-writer instead.
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。
基于 SOC 职业分类
完成 worktree 的完整合并流程:本地验证 → PR CI → merge → post-merge CI → 发布准备 → Release Notes → 创建 Release → 清理。 使用 git merge --no-ff 保留完整分支历史。 支持项目级钩子(.bare/custom-hooks/)实现个性化发布流程。 触发词:"合并worktree"、"merge-worktree"、"合并PR"、"发布"、"release"、"上线"。
Reference for writing workflow JS scripts. Auto-loaded when using workflow-generate or writing/editing workflow scripts for Pi. Covers runtime environment, injected globals, constraints, and script patterns. Not for general coding or subagent usage.
Use when the user says "轻量执行", "lite execute", "按 Wave 执行", "goal 模式执行 plan", "执行 plan", or has a completed plan.md (from lite-plan) or execution-plan.md (from mid-detail-plan) and needs to execute the Waves. ADR-029 后阶段 A+B 由 workflow run execute-full-workflow 机器接管(worktree-setup → dev waves → test+review → cleanup),主 agent 不再直接派 subagent。 对应 CW action: dev (workflow 内 implementer 渐进式提交) + test (workflow 内 test-runner 渐进式提交). Not for planning (lite-plan / mid-detail-plan). Not for retrospect (coding-retrospect).
Use when the user says "issue拆分", "问题分解", "issue decomposition", "方案对比", "tradeoff analysis", "优先级排序", or has finished system-architecture.md and needs to break it into prioritized issues with solution comparisons. Produces issues.md as a fog-of-war decision map. Step 3 of 6. Not for business requirements (Step 1) or architecture modeling (Step 2). Not for non-functional risk analysis (Step 4) or code-level design (Step 5).
Use when the user says "mid 详细计划", "中等功能实施设计", "issues+nfr+架构+计划一起做", "批量设计实施", or has finished mid-plan and needs issues.md + non-functional-design.md + code-architecture.md + execution-plan.md (+ code-skeleton) + detail.json (CW detail action 入参). 纯设计 skill,不写实现代码。对应 CW action: detail. Not for L3 heavy (use full-*). Not for requirements/architecture (that is mid-plan).
[internal] Shared reference files for the lite workflow (lite-plan / coding-execute / coding-retrospect). Not invoked directly — sibling lite-* skills resolve paths via ../lite-shared/references/{file}.md. Kept hidden from model invocation.
| name | meta-sk-agent-writer |
| description | Use when creating or editing agent.md files for Claude Code or Pi subagents. Not for SKILL.md — use meta-sk-skill-writer instead. |
agent.md 定义独立执行者。与 skill 不同,agent 有自己的上下文、工具集和模型选择,不继承父 agent 对话历史。
| 维度 | agent.md | SKILL.md |
|---|---|---|
| 触发 | 确定性(代码指定 agent name) | 概率匹配(模型看 description) |
| description | 能力摘要,不需要触发词 | 只写触发条件("Use when...") |
| 工具 | 可限制(tools/disallowedTools) | 无限制 |
| 模型 | 可指定(model 字段) | 继承主 agent |
| 上下文 | 独立,不继承父对话 | 在主 agent 上下文中 |
必填:
name: agent-name # 唯一标识符
description: "能力摘要" # 不需要触发词,写能力范围
可选(按使用频率排序):
| 字段 | 值 | 适用场景 |
|---|---|---|
model | inherit / sonnet / opus / haiku / 具体模型 ID | 审查用高质量模型,快速操作用轻量模型 |
tools | "Read, Bash, Grep" 或 ["Read", "Bash"] | 审查类只读,修复类需要 Edit |
skills | "skill-name" 或 ["s1", "s2"] | 启动时预加载 skill |
maxTurns | 正整数 | 限制 agent 轮次防止无限循环 |
memory | user / project / local | 跨会话持久记忆 |
background | true | 始终后台运行 |
effort | "high" / 正整数 | 推理努力级别 |
color | blue / green / yellow 等 | 终端颜色标识 |
omitClaudeMd | true | 只读 agent 省略 CLAUDE.md 注入(省 token) |
省略 tools = 所有工具可用。审查类 agent 建议限制为只读。
不需要"Use when..."格式。直接写能力范围:
# 好 — 能力摘要
description: "业务逻辑审查。验证代码覆盖所有业务用例,构造模拟数据推演执行路径。"
# 好 — 带边界
description: "代码品味审查(TS/Rust/Python)。读取品味文档后按 P0-P3 四级审查。不做 bug 检查。"
# 差 — 模糊
description: "审查代码。"
# 差 — 包含触发词(agent 不需要触发词)
description: "Use when you need BLR review."
# Agent 名称
一句话定位。
## 输入
task prompt 中必须/可选包含的参数。
## 执行步骤
1. 读取方法论(read skill 文件)
2. 获取数据
3. 执行分析
4. 输出结果
## 输出格式
具体格式定义。
## 约束
- 每条有后果链
agent 只定义"做什么"和"输出什么",方法论维护在 skill 中:
agent.md(~30-50 行)
→ read {skill_path}/SKILL.md(方法论)
修改方法论只需更新 skill,不改 agent。
源文件在项目中开发,symlink 到运行时目录:
| 工具 | agent 目录 |
|---|---|
| Claude Code | ~/.claude/agents/ |
| Pi | ~/.pi/agent/agents/ |
ln -s /path/to/agents/{name}.md ~/.claude/agents/{name}.md
ln -s /path/to/agents/{name}.md ~/.pi/agent/agents/{name}.md
| 字段 | Claude Code | Pi |
|---|---|---|
name / description | agent 发现和展示 | agent 发现和展示 |
model | 生效 | 不生效(Pi 用 model-resolve) |
tools | 生效 | 不生效(Pi 固定 read,bash,write,edit) |
| 正文 | system prompt 注入 | system prompt 注入 |
model 和 tools 在 Pi 中不生效,但保留不报错。正文在两个平台都作为 system prompt 注入。
meta-sk-skill-writermeta-sk-skill-writer/references/rule-templates.md