skills/write-output/SKILL.md
将数字员工的产出文件(产品规格文档、验收报告、设计文档等)写入共享工作区或个人工作区。 用于 PM 写 product_spec.md、Manager 写 review_result.md 等场景。 与 memory-save 的区别:本 Skill 专门用于写工作产出,memory-save 用于写记忆/会话状态。
npx skillsauth add kid0317/crewai_mas_demo write-outputInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
将数字员工完成的产出物(文档、报告等)持久化写入沙盒指定路径。
绝对禁止通过 sandbox_execute_bash 的 --content 参数传递文件内容:
# ❌ 已删除的旧方式(不要使用)
old_bash_approach(
cmd='python3 write_file.py --path "..." --content "大段内容..."' # shell 会截断内容!
)
原因:内容经过 JSON→shell 双重序列化,shell 遇到特殊字符会静默截断,
导致写入的文件内容不完整,但 errcode 仍可能返回 0(字节数是截断后的大小)。
必须使用 sandbox_file_operations 的 write action,内容通过 MCP 协议
以 JSON 字段直接传递,完全绕过 shell 解析,不存在特殊字符转义问题。
| 角色 | 产出路径 |
|------|----------|
| PM(产品规格) | /mnt/shared/design/product_spec.md |
| Manager(验收报告) | /workspace/review_result.md |
| Manager(需求文档) | /mnt/shared/needs/requirements.md |
| Dev(技术设计) | /workspace/tech_design.md |
sandbox_file_operations(action="write", ...) 实际写入文件expected_output 字段猜测结果调用 sandbox_file_operations 工具,参数如下:
action:"write"path:目标文件绝对路径content:完整文件内容(直接传入多行字符串,无需任何转义)示例:
sandbox_file_operations(
action="write",
path="/workspace/tech_design.md",
content="# 技术设计记录\n\n## T-01 ...\n\n### 接口定义\n..."
)
写入后立即用 sandbox_file_operations 读取,验证文件大小与内容:
sandbox_file_operations(
action="read",
path="/workspace/tech_design.md"
)
sandbox_file_operations(
action="write",
path="/mnt/shared/design/product_spec.md",
content="# 产品规格文档
## 产品概述
一句话描述产品
## 目标用户
...
## 用户故事
...
## 功能规格
F-01(P0): 用户注册
..."
)
sandbox_file_operations(
action="write",
path="/workspace/review_result.md",
content="# 验收报告
**验收结论**:✅ 通过
..."
)
sandbox_file_operations(
action="write",
path="/workspace/tech_design.md",
content="# 技术设计记录
**日期**: 2026-03-26
**类型**: tech_design
## T-01 自然语言日程解析模块技术设计
..."
)
如果 sandbox_file_operations 的 write action 报错,可用以下方式通过 Python 代码写入:
⚠️ 注意:content 内容中如包含三引号
"""字符,需将其替换为\"\"\",避免 Python 字符串提前终止。
sandbox_execute_code(
language="python",
code="""
content = \"\"\"(此处粘贴完整内容)\"\"\"
from pathlib import Path
p = Path("/workspace/tech_design.md")
p.parent.mkdir(parents=True, exist_ok=True)
p.write_text(content, encoding="utf-8")
print(f"写入成功:{p.stat().st_size} 字节")
"""
)
注意:不得使用 sandbox_execute_bash --content 方式,即便是备用方案也禁止。
content-media
技术设计文档 SOP——从需求到设计方案,产出 Markdown 文件到 /workspace/output/
content-media
技术设计文档 SOP——从需求到设计方案,产出 Markdown 文件到 /workspace/output/
content-media
技术设计文档 SOP——从需求到设计方案,产出 Markdown 文件到 /workspace/output/
development
Use this skill to persist important information from the conversation to workspace files so it survives across sessions. Activate proactively (without waiting for user to say "remember this") when: - User expresses a preference or habit ("I prefer...", "always...", "don't...") - User corrects Agent behavior and states how it should work instead - A key fact emerges that matters for future sessions (project milestone, decision made, important date, contact info) - User approves an approach ("let's do it this way going forward") - Agent needs to save a work product (report, spec, review result) to workspace Do NOT activate for: one-time tasks, Agent's own reasoning, info already in user.md.