ai

AIエージェント構築ガイド（LangChain/LangGraph・Google ADK・リアルタイムマルチモーダル）。 Use when building AI agents with LangChain, LangGraph, Google ADK, or Gemini Live API. フレームワーク選択、ツール定義、マルチエージェント、A2A、RAGパイプライン、リアルタイム音声/動画を含む。 For framework-agnostic GenAI design patterns, use designing-genai-patterns. For JS/Vercel AI SDK web integration, use integrating-ai-web-apps. For MCP protocol/server design, use lang:developing-mcp.

32 production-grade GenAI design patterns covering content control (Logits Masking, Grammar, Style Transfer), RAG architecture, model capabilities (Chain of Thought, Adapter Tuning), reliability (LLM-as-Judge, Reflection, Prompt Optimization), agentic systems (Tool Calling, Multiagent Collaboration), deployment optimization (SLM, Prompt Caching, Inference Optimization), safety guardrails (Self-Check, Guardrails), RAG implementation (11 source types, chunking, vector stores), and LLMOps/AgentOps (maturity L0–2, Tool/Agent Registry, Memory Governance, LLM metrics, LLMSecOps). Also covers AI performance (GPU/CUDA, LLM inference). Use when designing GenAI applications, choosing RAG strategies, implementing agent architectures, optimizing LLM reliability, or operating AI systems in production. For LangChain/LangGraph/ADK agents, use building-ai-agents. For web AI (Vercel AI SDK, LangChain.js), use integrating-ai-web-apps. For GCP-specific ML (BigQuery ML, Vertex AI), use cloud:developing-google-cloud.

AI-assisted development methodology covering prompt engineering, context engineering, code generation workflows, quality assurance, collaborative debugging, agent collaboration patterns, LLM dialogue design patterns (8 conversational patterns, Grammar of Intention, Test-Driven Prompting), team LLM adoption workshops, and role-specific augmentation (Dev/PO/Coach/Manager). MUST load when working with AI coding tools or discussing AI-assisted development practices. Use when crafting prompts for code generation, managing AI context, reviewing AI-generated code, coordinating multi-agent workflows, or designing structured LLM dialogues.

LLM evaluation and red-teaming toolkit using promptfoo. Covers promptfooconfig.yaml configuration, 40+ assertion types (deterministic, model-graded, RAG), provider setup (OpenAI, Anthropic, Google, Ollama, HTTP, custom JS/Python), red teaming (134+ plugins, jailbreak strategies, compliance frameworks), CLI commands, caching, and CI/CD integration. Use when writing promptfooconfig.yaml, designing LLM test suites, running adversarial red team evaluations, or integrating LLM quality gates in CI/CD. Detects: promptfooconfig.yaml or promptfoo in package.json. For general LLMOps operations, use designing-genai-patterns. For general test methodology (TDD/AAA), use devkit:testing-code.

Comprehensive guide to integrating generative AI into web applications using Vercel AI SDK, LangChain.js, and MCP with React/Next.js. Covers streaming chat, RAG and summarization, tool calling, structured data generation, prompt engineering, AI app testing (mocking, E2E, evaluation), and deployment security (API key management, CORS, rate limiting). Use when building AI-powered web features in JavaScript/TypeScript projects. For framework-agnostic GenAI design patterns and RAG system internals (vector stores, chunking, LLMOps), use designing-genai-patterns. For MCP protocol/server development, use lang:developing-mcp. For Next.js framework specifics, use web:developing-nextjs. For Python LangChain/LangGraph agent development, use building-ai-agents. For Python/FastAPI generative-AI service backends (model serving, streaming), use web:developing-fastapi.

Organizational security strategy for AI-powered software development covering trust frameworks (Early/Explainable/Experience), trust platform maturity model, adaptive governance with dynamic policies, AI-BOM, AI-SPM, cross-functional ownership, risk landscape (CWE vulnerability patterns in AI-generated code, hallucinated package references, OWASP Agentic AI threat categories), security KPIs/metrics, secure prompt templates, incident response updates, and a phased implementation roadmap. Use when establishing security controls, governance programs, or usage policies for AI coding assistants, agentic systems, or AI-accelerated SDLC workflows. For code-level security (OWASP, CodeGuard), use devkit:securing-code instead. For AI development methodology (prompts, context engineering), use developing-with-ai instead. For LLM-specific security (prompt injection, LLMSecOps), use designing-genai-patterns.