⌘K
Change language Switch ThemeSign In
Narrow Mode
Google Cloud 总结：每位 ADK 开发者都该知道的 5 种 Agent Skill 设计模式
山 山行AI @山行AI
One Sentence Summary
This article organizes the 5 ADK Agent Skill design patterns summarized by Google Cloud: Tool Wrapper, Generator, Reviewer, Inversion, and Pipeline, each suitable for scenarios like knowledge injection, structured output, standardized review, requirement clarification, and process control, with guidance on pattern selection and combination recommendations.
Summary
Based on Google Cloud's official content, this article systematically organizes 5 high-frequency Skill design patterns in ADK Agent development. The Tool Wrapper pattern is used to load specific tech stack knowledge on demand, avoiding context pollution; the Generator pattern achieves stable structured output through templates; the Reviewer pattern modularizes review rules for reusable checking processes; the Inversion pattern reverses the workflow, letting the Agent first ask questions to clarify requirements before acting; the Pipeline pattern ensures execution quality for complex tasks through serial steps and human confirmation nodes. The article provides the applicable scenarios, implementation key points, and a pattern selection decision framework for each pattern, emphasizing that these patterns can be combined to build more powerful Agent structures. This is a practical AI Agent development methodology guide.
Main Points
* 1. Skill design patterns are more important than format specifications; specifications solve 'how to package' a Skill, while patterns solve 'how to design the soul of a Skill'ADK's SKILL.md format specification only solves the file structure problem, while what truly affects Agent effectiveness is the internal logic design of Skills. The 5 patterns summarized by Google Cloud fill this gap. * 2. Tool Wrapper pattern achieves on-demand loading of specific tech stack knowledge through keyword monitoring, avoiding system prompt bloatThis pattern doesn't inject a large set of rules all at once, but monitors tech stack keywords in user questions and loads corresponding reference materials only when命中 (hit/matched), making the Agent a 'temporary expert' that switches on demand. * 3. Reviewer pattern modularizes review standards into independent checklists, enabling multi-scenario reuse of the same structureBy placing review rules in a separate checklist file, the same Reviewer structure can quickly switch between code reviewer, security reviewer, or API design reviewer. * 4. Inversion pattern reverses the Agent's working mode, transforming it from 'immediately giving answers' to 'an interviewer who first builds a problem model'Through strong constraints requiring the Agent to first ask clarifying questions in ambiguous requirement scenarios, including user type, deployment environment, constraints, etc., to avoid direction errors from premature action. * 5. The 5 patterns can be combined: Pipeline can end with Reviewer for quality review, and Generator's first step can connect to Inversion for variable clarificationPatterns themselves are not the goal; maintainable, verifiable, and reusable Agent structures are the goal, and combining them creates greater value.
Metadata
AI Score
81
Website mp.weixin.qq.com
Published At Today
Length 2102 words (about 9 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
山行AI 2026-03-20 19:28 浙江
Google Cloud 总结：每位 ADK 开发者都该知道的 5 种 Agent Skill 设计模式
> 原始链接：https://x.com/GoogleCloudTech/status/2033953579824758855[1] > > > 原作者：Google Cloud Tech，@Saboo_Shubham_，@lavinigam > > > 说明：以下内容为中文整理与翻译版，配图来自原帖。
当越来越多的 Agent 工具开始采用统一的 SKILL.md 结构之后，真正拉开差距的已经不再是“文件格式写得对不对”，而是“技能内容该怎么设计”。
原文的核心观点很清晰：规范只告诉你如何打包一个 Skill，却没有告诉你怎样设计 Skill 内部的逻辑。而在真实开发里，一个封装 FastAPI 规范的 Skill，和一个分 4 步执行的文档流水线 Skill，虽然外表都叫 SKILL.md，但内部结构完全不是一回事。
Google Cloud 将这些常见做法总结成 5 种高频设计模式：
* Tool Wrapper：把某个库或框架的知识即时注入给 Agent
* Generator：按固定模板生成结构化输出
* Reviewer：按清单逐项审查并给出分级结论
* Inversion：先反过来提问，再开始行动
* Pipeline：强制 Agent 按顺序执行复杂流程
一、Tool Wrapper：让 Agent 立刻成为某个技术栈的“领域专家”
Tool Wrapper 的目标，是让 Agent 在处理某个特定技术栈时，按需加载该栈的内部规范，而不是把大量约定硬塞进系统提示词。这种模式最适合：
* 团队内部编码规范分发
* 特定框架最佳实践注入
* 库级 API 使用约束
* 让 Agent 只在“需要时”加载上下文
它的重点不在于“写很多规则”，而在于：
* 监听用户提问中的特定关键词
* 在命中相关技术栈时，再去加载 references/ 中的资料
* 将这些资料视为该场景下的最高优先级规则
这样一来，Agent 平时不会被无关知识污染，只有在真正处理 FastAPI、React、Pydantic 等问题时，才临时变成该领域的专家。
二、Generator：把“每次都不一样”变成“稳定输出”
如果你经常遇到 Agent 每次生成的文档结构都不同，那么 Generator 模式就是最直接的解决方案。
它的思路是：
* 把输出结构放到 assets/ 里
* 把语气、格式、风格要求放到 references/ 里
* Skill 本身只负责协调执行顺序
这种模式特别适合：
* 技术报告生成
* API 文档模板化输出
* 周报、总结、方案文档统一格式
* 项目脚手架类文本产物生成
它本质上是在让 Agent 扮演“模板执行器”而不是“自由发挥的写手”。
你不再要求它每次都重新思考文档结构，而是要求它先读模板、再补变量、最后按风格指南填完整份文档。
三、Reviewer：把“怎么检查”与“检查什么”拆开
Reviewer 模式的核心，是把审查规则模块化。传统做法往往把所有代码异味、风格要求、安全要求都写进一大段提示词里，结果既难维护，也很难复用。
而 Reviewer 模式会把检查标准独立放进一个清单，比如：
* references/review-checklist.md
然后让 Agent：
- 先理解代码在做什么
- 再按清单一项项审查
- 对每个问题标记严重级别
- 给出原因与修复建议
* PR 自动审查
* 安全清单扫描
* 风格检查
* 质量审计
最大的好处是：
你只需要替换“检查清单”，就能把同一套 Reviewer 结构，切换成 Python 代码审查器、OWASP 安全审查器，或者 API 设计审查器。
四、Inversion：让 Agent 先采访你，而不是先猜
Agent 天生倾向于“尽快给答案”，但很多场景下，这反而会让它在信息不完整时过早行动。
Inversion 模式就是反过来设计：先提问，后输出。
它通常会加入强约束，例如：
* 在所有阶段完成前，不允许直接开始设计
* 一次只问一个问题
* 必须把需求、约束、用户类型、部署环境都问清楚
这种模式最适合：
* 新项目规划
* 系统设计访谈
* 需求不明确的复杂任务
* 高风险方案生成
它的价值在于：
让 Agent 从“立刻给结论的执行者”，变成“先建立问题模型的采访者”。
五、Pipeline：复杂任务必须按步骤走完，不能跳
当一个任务步骤很多、并且中间结果必须人工确认时，就需要 Pipeline 模式。
这种模式强调：
* 步骤必须串行
* 每一步都有明确输入与输出
* 失败不能跳过
* 关键节点需要人工确认后才能继续
常见适用场景：
* 文档流水线
* 多阶段发布流程
* 审核制生成流程
* 需要阶段性验收的复杂任务
它本质上是把 Skill 写成一个“带关卡的工作流”，而不是一段自由提示词。
Agent 不能直接跳到最后一步给你一个看起来完整、实际上未经验证的结果。
六、如何选型：先问自己“你真正缺的是什么”
原文给出的判断思路可以归纳成一句话：
* 如果你缺的是“特定知识注入”，选 Tool Wrapper
* 如果你缺的是“固定输出结构”，选 Generator
* 如果你缺的是“标准化审查”，选 Reviewer
* 如果你缺的是“需求澄清”，选 Inversion
* 如果你缺的是“严格流程控制”，选 Pipeline
很多团队一开始就把所有要求塞进一个超长系统提示词，结果提示词越来越脆、越来越难维护。
更好的方式，是把任务拆开，然后针对不同问题选择最合适的结构模式。
七、这些模式不是互斥的，真正强大的是“组合”
这 5 种模式不是“只能选一个”，而是可以组合使用：
* Pipeline 的最后一步可以接一个 Reviewer，自动做质量复核
* Generator 的第一步可以接 Inversion，先把变量问清楚
* Tool Wrapper 可以被嵌入进更复杂的流程节点里
换句话说，模式本身不是目的，可维护、可验证、可复用的 Agent 结构才是目的。
八、结语
这篇内容最值得记住的一点，不是 5 个模式的名字，而是它背后的设计观：
不要再试图把复杂、脆弱、彼此冲突的指令都塞进一条系统提示词。
更稳妥的做法，是把任务拆成结构清晰的模式，再让 Agent 在运行时按需加载对应的上下文与步骤。
格式规范解决的是“怎么打包”。
而这 5 种模式解决的是“怎么设计 Skill 的灵魂”。
原文来源
* Google Cloud Tech on X: https://x.com/GoogleCloudTech/status/2033953579824758855[2]
引用链接
[1]_https://x.com/GoogleCloudTech/status/2033953579824758855_
[2]_https://x.com/GoogleCloudTech/status/2033953579824758855_ 跳转微信打开
山 山行AI @山行AI
One Sentence Summary
This article organizes the 5 ADK Agent Skill design patterns summarized by Google Cloud: Tool Wrapper, Generator, Reviewer, Inversion, and Pipeline, each suitable for scenarios like knowledge injection, structured output, standardized review, requirement clarification, and process control, with guidance on pattern selection and combination recommendations.
Summary
Based on Google Cloud's official content, this article systematically organizes 5 high-frequency Skill design patterns in ADK Agent development. The Tool Wrapper pattern is used to load specific tech stack knowledge on demand, avoiding context pollution; the Generator pattern achieves stable structured output through templates; the Reviewer pattern modularizes review rules for reusable checking processes; the Inversion pattern reverses the workflow, letting the Agent first ask questions to clarify requirements before acting; the Pipeline pattern ensures execution quality for complex tasks through serial steps and human confirmation nodes. The article provides the applicable scenarios, implementation key points, and a pattern selection decision framework for each pattern, emphasizing that these patterns can be combined to build more powerful Agent structures. This is a practical AI Agent development methodology guide.
Main Points
* 1. Skill design patterns are more important than format specifications; specifications solve 'how to package' a Skill, while patterns solve 'how to design the soul of a Skill'
ADK's SKILL.md format specification only solves the file structure problem, while what truly affects Agent effectiveness is the internal logic design of Skills. The 5 patterns summarized by Google Cloud fill this gap.
* 2. Tool Wrapper pattern achieves on-demand loading of specific tech stack knowledge through keyword monitoring, avoiding system prompt bloat
This pattern doesn't inject a large set of rules all at once, but monitors tech stack keywords in user questions and loads corresponding reference materials only when命中 (hit/matched), making the Agent a 'temporary expert' that switches on demand.
* 3. Reviewer pattern modularizes review standards into independent checklists, enabling multi-scenario reuse of the same structure
By placing review rules in a separate checklist file, the same Reviewer structure can quickly switch between code reviewer, security reviewer, or API design reviewer.
* 4. Inversion pattern reverses the Agent's working mode, transforming it from 'immediately giving answers' to 'an interviewer who first builds a problem model'
Through strong constraints requiring the Agent to first ask clarifying questions in ambiguous requirement scenarios, including user type, deployment environment, constraints, etc., to avoid direction errors from premature action.
* 5. The 5 patterns can be combined: Pipeline can end with Reviewer for quality review, and Generator's first step can connect to Inversion for variable clarification
Patterns themselves are not the goal; maintainable, verifiable, and reusable Agent structures are the goal, and combining them creates greater value.
Key Quotes
* Specifications only tell you how to package a Skill, but not how to design the internal logic of a Skill * The Agent won't be polluted by irrelevant knowledge under normal circumstances; it only temporarily becomes an expert in that field when actually handling issues with FastAPI, React, Pydantic, etc. * It essentially makes the Agent act as a 'template executor' rather than a 'freelance writer' * You only need to replace the 'checklist' to switch the same Reviewer structure into a Python code reviewer, OWASP security reviewer, or API design reviewer * Transforming the Agent from 'an executor who immediately gives conclusions' to 'an interviewer who first builds a problem model' * Stop trying to stuff complex, fragile, and mutually conflicting instructions into a single system prompt
AI Score
81
Website mp.weixin.qq.com
Published At Today
Length 2102 words (about 9 min)
Tags
ADK
Agent Skill
Design Patterns
Google Cloud
AI Agent
Related Articles
* 128. Manus Co-founder's Final Interview Before Sale: The Fantastical Drift to 2025... * Two Open Source "AI Cowork" Desktop Tools - Eigent and AionUi * GPT-5.4 Released: OpenAI's First Unified Model, Truly Native * Deconstructing Clawdbot: Local Architecture, Memory Management, Agent Orchestration, and Context Assembly Principles * LangExtract: An LLM-Powered Tool for Structured Extraction * 1000x Productivity Revolution: Claude Code Creator Reveals the 'Bitter Lesson' of AI Development and the Future Truth * How to Design an AI Agent System * CLI-Anything: Making Any Software Agent-Driven * What You Don't Know About Claude Code: Architecture, Governance, and Engineering Practices - Tw93 * The Era of Vibe Coding: How to Break In? HomeArticlesPodcastsVideosTweets