AgentSkillsCN

recipes

为 AI 辅助工作流提供可复用的开发模式与指南。适用于询问代理的 CLI 设计、注释标准、快照测试、凭证存储、Python 项目架构、语义缩放、技能编写、Claude 工具、提取确定性工作,或插件的开发与分发时使用。

SKILL.md
--- frontmatter
name: recipes
description: Reusable development patterns and guides for AI-assisted workflows. Use when asking about CLI design for agents, commenting standards, snapshot testing, credential storage, Python project architecture, semantic zoom, skill writing, Claude tools, extracting deterministic work, or plugin authoring and distribution.

<essential_principles>

What Recipes Are

Recipes are standalone reference documents -- each captures a proven pattern or practice for AI-assisted development. They are not step-by-step procedures; they are knowledge you read and apply.

Three categories:

CategoryRecipesWhen to Use
AI Agent Patternssemantic-zoom, agentic-cli, extract-deterministicDesigning AI interactions or CLI tools for agents
Development Practicescomments, agent-skills, claude-tools, snapshot-testingWriting code, skills, or tests in this ecosystem
Distributionclaude-plugin-authoringPackaging skills as installable Claude Code plugins
Architecturepython-project-architecture, keyring-credential-storageStructuring Python projects or handling credentials

</essential_principles>

<intake> What pattern or practice do you need guidance on?
  1. AI agent CLI design - Safe-by-default CLI patterns for agentic workflows
  2. Extracting deterministic work - When to script vs. when to prompt
  3. Semantic zoom - Controlling abstraction level in AI interactions
  4. Python commenting standards - INTENT:, CRITICAL:, PERF: anchors
  5. Writing Claude Code skills - Skill authoring patterns and best practices
  6. Claude Code tools reference - Task tool, subagents, tool usage
  7. Snapshot testing - Approval testing with syrupy
  8. Python project architecture - Project structure for CLI tools
  9. Keyring credential storage - Secure credential management patterns
  10. Claude plugin authoring - Packaging skills as installable plugins

Wait for response before proceeding. </intake>

<routing>
ResponseReference
1, "cli", "agentic", "safe", "flags", "non-interactive"references/agentic-cli.md
2, "deterministic", "extract", "script vs prompt"references/extract-deterministic.md
3, "zoom", "abstraction", "detail level", "semantic"references/semantic-zoom.md
4, "comment", "intent", "critical", "perf", "anchor"references/comments.md
5, "skill", "writing", "skill.md", "author"references/agent-skills.md
6, "tools", "task tool", "subagent", "claude tools"references/claude-tools.md
7, "snapshot", "testing", "syrupy", "approval"references/snapshot-testing.md
8, "architecture", "python project", "cli tool", "structure"references/python-project-architecture.md
9, "keyring", "credential", "password", "secret"references/keyring-credential-storage.md
10, "plugin", "marketplace", "distribute", "package", "install"references/claude-plugin-authoring.md
OtherClarify intent, then select appropriate reference

After identifying the reference, read it and apply its guidance to the user's situation.

</routing>

<reference_index>

AI Agent Patterns

  • agentic-cli.md - CLI design patterns for AI agents (non-interactive, capability gradient, dry-run)
  • extract-deterministic.md - When to extract deterministic work from AI workflows into scripts
  • semantic-zoom.md - Controlling abstraction level in AI interactions

Development Practices

Distribution

Architecture & Implementation

</reference_index>

<success_criteria>

  • User gets directed to the right reference for their question
  • Reference content is applied to the user's specific situation, not just recited
  • Multiple references can be combined when a question spans categories

</success_criteria>