LLM 與 Agent

Research external APIs/services/libraries for feature implementation. Interviews user, researches options, confirms choice, gathers implementation notes. Creates .claude/plans file.

Grep/wc replacement for .runs artifacts. Use for: count, extract, Machine Summary, receipt reading, marker counts. Null-safe counting (REQ/NFR/QID/RSK markers), YAML block parsing, BDD scenario counting. Deterministic read-only - no judgment. Use when cleanup agents need mechanical counts/extraction. Invoke via bash .claude/scripts/demoswarm.sh.

Google Gemini API 集成指南：提供 AI 分析、提示词工程、流式响应和多模态处理的最佳实践。Use when: 需要集成 Gemini API、编写提示词、处理流式响应、多模态输入。Triggers: "Gemini", "AI", "LLM", "提示词", "prompt", "流式", "streaming", "多模态"

HestAI workflow phase execution including D0→D1→D2→D3→B0→B1→B2→B3→B4→B5 sequence, agent assignments per phase, phase transition validation, and deliverable requirements. Use when starting new phase, determining next steps, or validating phase completion.

AI-powered pipeline diagnostics. Explains why faces weren't detected, tracked, embedded, or clustered using OpenAI GPT-4o analysis. Use when debugging pipeline issues, reviewing unclustered tracks, or understanding why quality gates rejected faces.

This skill should be used when the user asks about "MakerKit patterns", "CLAUDE.md files", "what MCP tools are available", "how does MakerKit structure", "personal vs team accounts", "RLS patterns in MakerKit", or needs quick reference for MakerKit conventions.

Comprehensive audit of Claude Code hooks for correctness, safety, and performance. Use when reviewing, validating, or debugging hooks, checking JSON stdin handling, verifying exit codes (0=allow, 2=block), analyzing error handling, fixing hook failures, ensuring safe degradation, optimizing performance, or validating settings.json registration. Also triggers when user asks about hook best practices, wants to create a new hook, or needs help with hook configuration.

Build stateful AI agents and agentic workflows with LangGraph in Python. Covers tool-using agents with LLM-tool loops, branching workflows, conversation memory, human-in-the-loop oversight, and production monitoring. Use when: (1) building agents that use tools and loop until task complete, (2) creating multi-step workflows with conditional branches, (3) adding persistence/memory across turns with checkpointers, (4) implementing human approval with interrupt(), (5) debugging via time-travel or LangSmith. Covers StateGraph, nodes, edges, add_conditional_edges, MessagesState, thread_id, Command objects, and ToolMessage handling. Examples include chatbots, calculator agents, and structured workflows.

Automatically fix ESLint errors by modifying code to comply with linting rules. For small codebases (≤20 errors), fixes directly. For larger codebases (>20 errors), spawns parallel agents per directory for efficient processing. Never disables rules or adds ignore comments.

FOLLOW THE STATE MACHINE IN SKILL.MD. When user says 'continue': (1) FIRST: Run pwd, (2) Announce STATE: CHECK_STATUS, (3) Read .claude/session.md to check Status field, (4) Route based on Status. NEVER auto-advance tasks. NEVER use TodoWrite. NEVER create git commits.

Guide for creating effective skills. This skill should be used when users want to create a new skill (or update an existing skill) that extends Claude's capabilities with specialized knowledge, workflows, or tool integrations.

Review code changes for correctness, security, performance, and maintainability. Use for PR reviews,code audits, pre-merge checks, or quality validation of Laravel + React code. EXCLUSIVE to reviewer agent.

Create custom Claude Code subagents with proper structure and best practices. Supports project (.claude/agents/) and global (~/.claude/agents/) scopes. Activated when users mention "create agent", "new subagent", "add agent", "agent 만들기", or need to build specialized AI assistants.

dotfileの編集・修正を依頼されたとき、ホームディレクトリのファイルではなくchezmoiで管理されているファイルを優先的に確認・編集する。対象：~/.bashrc、~/.zshrc、~/.vimrc、~/.rsync-exclude、~/.config/**、~/.ssh/**、~/.claude/**など、ホームディレクトリの`.`で始まるファイル・フォルダ配下の全ファイル。

Compile visual workflow canvas to Google ADK Python code. Use when generating executable agent code from WorkflowCanvas data structures. Supports all ADK workflow patterns.

Guide for creating effective skills. This skill should be used when users want to create a new skill (or update an existing skill) that extends Claude's capabilities with specialized knowledge, workflows, or tool integrations.

Expert assistant in free and open-source licenses (FOSS). Helps users choose the most suitable free/open license for their creations through a guided questionnaire. Use when user asks about open-source licensing, GPL, MIT, Apache, Creative Commons, copyleft, permissive licenses, or wants to make their project open-source. Covers software licenses, Creative Commons for creative works, and region-specific licenses like CeCILL. Responds in the user's language.

Comprehensive guide to Claude Code's extensibility features including subagents, skills, plugins, commands, hooks, MCP servers, permissions, and automation. Use when the user asks about Claude Code features, capabilities, customization, extensibility, how to extend Claude Code, or automation workflows.

Use after planning or implementing non-trivial tasks - runs Codex MCP background verification/review for quality gates (plan validation + implementation review)

Generate comprehensive coverage report with requirement traceability mapping. Shows coverage per REQ-*, gaps, trends, and recommendations. Use for status dashboards, quality reviews, or compliance audits.