5 Proven Agent Skill Design Patterns Google Shares After Anthropic

Tool Wrapper: Let the agent quickly become an expert in a domain

This pattern packages a library or framework’s specification documents into a skill. The agent loads the relevant documentation only when it detects the corresponding library keyword, such as FastAPI, and then treats the loaded conventions.md as authoritative rules.

Generator: Produce Structured Output from Templates

Generator enforces a consistent output format by using two optional directories: assets/ for output templates and references/ for style guides. The skill acts as a project manager, directing the agent to load a template, read the style guide, ask the user for missing variables, and finally fill the document. This is useful for generating uniform API docs, standardized commit messages, or scaffolded project structures.

Reviewer: Separate Checklist from Review Logic

Reviewer decouples “what to check” from “how to check.” The checklist (e.g., references/review-checklist.md) can contain Python‑style checks, OWASP security checks, or any other criteria. By swapping the checklist file, the same skill infrastructure can perform completely different audits without changing the core logic.

Inversion: Agent Asks Questions Before Acting

Inversion flips the usual flow: the agent first asks a series of non‑negotiable gating questions (e.g., “Do not start building until all stages are completed”). It waits for the user’s answers before proceeding to the next stage, ensuring that required information is collected up front.

Pipeline: Strict Workflow with Hard Checkpoints

Pipeline enforces a rigid sequence of steps with explicit gate conditions. For example, a document pipeline may consist of parsing, generating a document string, assembling the document, and quality checking. Each step requires user confirmation before the agent can move to the next, preventing the agent from skipping verification.

Choosing the Right Pattern

Each pattern fits specific scenarios; the guide includes a decision diagram to help select the appropriate pattern based on task complexity and required control.

Patterns Can Be Combined

The five patterns are not mutually exclusive. A Pipeline can end with a Reviewer step for double‑checking, while a Generator can start with an Inversion step to gather necessary variables. ADK’s SkillToolset and progressive disclosure ensure that only the needed pattern loads at runtime, conserving context tokens.

Anthropic discovered that the best Skills are not finely tuned prompts but a “toolbox.” They categorize Skills into nine types, each with clear scenarios. Their three iron rules for good Skills are: write only what the agent doesn’t already know, focus on pitfall checklists, and give tools—not instructions. Reference: Anthropic engineers share Claude Code skill design guide – 9 types and practical tips https://wink.run/pings/content/111668?from=wx

Related links:

Original article: https://x.com/i/article/2033941492633362432

Awesome Agent Skills repository: https://github.com/skillmatic-ai/awesome-agent-skills