AgentSkillsCN

agent-authoring

如何为本工作空间设计、构建并验证自定义 Copilot 代理。

SKILL.md
--- frontmatter
name: agent-authoring
description: How to design, structure, and validate custom Copilot agents for this workspace.

Agent Authoring Skill

Guidelines for creating and maintaining custom GitHub Copilot agents (.agent.md) tailored to this repository.

When to Use

  • •You need a new Copilot agent focused on a workflow (e.g., Storybook tests, log triage).
  • •You want to adjust tools/persona for existing agents.
  • •You are adding workspace-level agents under .github/agents or profile-level agents for reuse.

Quick Start

  1. •Create .github/agents/<agent-name>.agent.md (VS Code auto-detects this folder/extension).
  2. •Add YAML frontmatter with only supported fields: name, description, tools, target, infer, mcp-servers, handoffs.
  3. •In the body, write concise instructions: task focus, guardrails, run/validate steps, and handoffs.
  4. •Keep it under ~200-300 lines; link out to docs instead of pasting everything.

Frontmatter Cheatsheet

  • •name: Display name; defaults to file name if omitted.
  • •description: Placeholder text in chat input; keep short.
  • •tools: List allowed tools. For MCP servers, use <server>/* to include all tools or enumerate specific ones.
  • •target: github-copilot for Copilot agents (default), vscode for VS Code target.
  • •infer: Allow subagent use (default true). Set false if you want to restrict subagents.
  • •mcp-servers: For Copilot agents, point to MCP server configs (paths to server JSON). Use only if you need to pin servers.
  • •handoffs: Optional array of next-step buttons { label, agent, prompt?, send? }.

Body Structure (recommended)

  • •Purpose and scope: What this agent does and what it must not do.
  • •Operating defaults: Where to run commands, what to inspect first, preferred scripts.
  • •Playbooks: Ordered steps for common flows (run, debug, fix, validate).
  • •Guardrails: Safe-edit rules (scope, no unrelated changes, ask before destructive ops).
  • •Validation: How to verify work (tests/linters), how to summarize outputs.
  • •Links: Point to repo docs or skills for deeper detail.

Tool Selection Patterns

  • •Editing agent: include workspace, terminal, search; add fetch for docs; add MCP toolsets only if required.
  • •Read-only agent: omit editing tools; keep fetch/search.
  • •Narrow the tool list to prevent unsafe actions (e.g., no terminal for planning-only agents).

Handoffs

Use handoffs to guide sequential flows (e.g., Plan -> Implement -> Review). Keep labels short and make prompts actionable. Use send: false unless you want auto-submit.

Quality Bar

  • •Minimal, task-focused instructions; avoid boilerplate.
  • •No unsupported frontmatter fields (reject argument-hint, etc.).
  • •Prefer scoped commands over global installs; mention working directory if non-root.
  • •Add guardrails for destructive commands (never propose git reset --hard).
  • •Keep instructions ASCII and concise; use fenced code blocks for commands.

Validation Checklist

  • • Supported frontmatter only; fields spelled correctly.
  • • Tools list matches the agent’s intent; no over-broad capabilities.
  • • Clear run/validate steps with default commands.
  • • Guardrails called out.
  • • Links resolve (relative paths within repo or official docs).

Examples

  • •Location: .github/agents/storybook-interaction-tests.agent.md (interaction test fixer playbook).
  • •Planning agent: frontmatter with tools: ['fetch', 'search'], body instructs to gather context and produce a plan, handoff to implementation agent.

Maintenance Tips

  • •Keep agents in sync with repo workflows (scripts, test commands). Update after script changes.
  • •Remove stale handoffs when target agents are renamed or removed.
  • •Prefer additive changes; avoid reformatting existing agents without need.