菜单
Author a new ZLAgent skill in the canonical Trigger / Inputs / Steps / Verification / Failure Signals format. Class-level only. Used when the user says "把这个流程记成 skill" or in background skill review.
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
| name | skill-authoring |
| description | Author a new ZLAgent skill in the canonical Trigger / Inputs / Steps / Verification / Failure Signals format. Class-level only. Used when the user says "把这个流程记成 skill" or in background skill review. |
| version | 0.1.0 |
| tags | ["meta","skills"] |
| metadata | {"zlagent":{"category":"meta","triggers":["记成 skill","写一个技能","保存为 skill","author skill","new skill"],"capabilities":[],"related_skills":["task-planning","systematic-debugging"]}} |
一切 skill 都应当遵循钱学森《论系统工程》的"目标—输入—执行—输出—反馈" 控制论闭环。本 skill 教 agent 把这个框架套进 SKILL.md， 模板原型见
workspace/skills/_TEMPLATE.md。
把一个类级别（class-level，至少 2 次以上可复用）的工作流固化成 SKILL.md，让未来同类任务可以一次命中、一次执行、一次校验。
不属于本 skill 的范围：
skill_manage action='patch'正向触发：
反向排除：见上文"目标"小节末尾的边界条件。
name: <kebab-case-name>
description: |
一句话说明触发场景 + 可观测产出。≤200 字符。
version: 0.1.0
tags: [<category>, ...]
metadata:
zlagent:
category: <category>
triggers: [...]
capabilities: []
related_skills: [...]
## 一、目标 / Goal## 二、触发 / Trigger（输入信号）## 三、步骤 / Steps（执行控制）## 四、输出 / Output（被控变量）## 五、反馈 / Feedback（偏差检测 + 校正）references/。skill_manage(action='create', ...)。
skill_body 不要包含外层 YAML --- 分隔符——skill_manage 自己渲染。涉及代码 / 调试 / 重构 / 实现方案时遵循 Karpathy-inspired 原则： 编码前思考、简洁优先、精准修改、目标驱动执行。
skill_manage 创建成功的 skill 应当满足：
name + description。tool_search）。偏差信号——出现任一项就立即停手，不要把这种 skill 写进 workspace：
references/patch 而不是 create校正策略：
skill_manage(action='patch', ...)workspace/skills/_TEMPLATE.md 重新对齐Fetch and summarise recent papers from arxiv.org. Use this skill when the user wants a digest of today's or this week's cs.AI / cs.LG / cs.CL listings, a focused per-topic roundup, or a quick look at a specific submission.
Lightweight code-review checklist the agent runs before declaring a patch ready. Catches the obvious classes of issue (dead code, missing error handling, weakened tests) without a full security audit. Adapted from addyosmani's "Review" lifecycle phase.
Root-cause-first debugging loop for code or runtime errors: understand the failure before proposing a fix. Adapted from Hermes' systematic- debugging — same iron law, smaller surface, ZLAgent-tool-aware.
Do not declare a task done without running verification. For code, that means running tests / smoke; for answers, that means checking the claim against a fresh source. Adapted from obra/superpowers and addyosmani's "Verify" lifecycle phase.
Umbrella skill for user document workflows — PDF, Word, Excel, PowerPoint, text extraction, summarization. Dispatches to the right MCP / tool; proposes an install via mcp-discovery when no local capability fits. Adapted from Anthropic's docx/pdf/xlsx/pptx skills.
When the user wants something the agent doesn't have a tool for yet (travel planning, GitHub issues, Linear, weather, filesystem, ...), search the live MCP ecosystem for an existing server, propose installing it, and on confirm wire it up via mcp_manage. Adapted from Hermes' mcporter skill — same flow, ZLAgent-tool-aware.