LLM 與 Agent

タスクの複雑さに応じて Claude のモデルを自動的に切り替える。複雑なタスク（アーキテクチャ設計、大規模リファクタリング、難解なバグ修正）では opus を、単純なタスク（小さな修正、フォーマット、簡単な質問）では sonnet を自動選択。コストと品質のバランスを最適化。

Local RAG with Ollama and FAISS

Use when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies

Systematic technical debt reduction and code optimization agent. Use when refactoring code, reducing complexity, eliminating code smells, improving performance, cleaning up unused code, or modernizing legacy patterns. Handles dependency updates, architecture improvements, and codebase health metrics.

Autonomous development workflow. Generate detailed specs, plans, and tasks for autonomous agent execution with session memory tracking.

Skills for working with the agent-plane project itself. Demonstrates how skills work.

Complete visual QA - screenshots with full-page scroll to trigger animations, then analyze with Claude. Run after CSS/template changes.

Transform messy code into clean code through disciplined refactoring. Code smells, refactoring patterns, and the courage to improve without breaking. Leave every codebase better than you found it.

Guide for implementing Cloudflare R2 - S3-compatible object storage with zero egress fees. Use when implementing file storage, uploads/downloads, data migration to/from R2, configuring buckets, integrating with Workers, or working with R2 APIs and SDKs.

Generates comprehensive BDD/Gherkin tests for features. Triggers when user needs to write tests, create test scenarios, add test coverage, or verify feature completeness.

Design reusable prompt templates that encode domain-specific patterns for recurring AI development tasks.Use this skill when you've executed similar prompts 2+ times and need to capture the pattern as reusableintelligence. Transforms one-off prompts into parameterized templates that maintain quality while reducingcognitive load. Helps students move from Layer 2 (AI Collaboration) to Layer 3 (Intelligence Design) inthe 4-Layer Teaching Method.

Official Anthropic SDK for Claude AI with chat, streaming, function calling, and vision capabilities

Agent coordination, orchestration, and multi-agent workflow management scripts

Stack prompts in queue while agent runs, minimizing context switches to achieve sustained 6k lines per day output and preserved flow state. Use for heavy coding sessions, iterative refinements, rapid iteration cycles, or sustained productivity. Queue enables async execution without blocking. Triggers on "flow state", "queue prompts", "heavy coding", "sustained output", "minimize interruptions".

Set up and orchestrate multi-agent workflows using Beads (shared issue tracking/memory) and MCP Agent Mail (agent messaging). Use when coordinating multiple agents, setting up agent villages, tracking work across agent sessions, or implementing swarm workflows.

Production-ready starter project for React + Cloudflare Workers + Hono with core services(D1, KV, R2, Workers AI) and optional advanced features (Clerk Auth, AI Chat, Queues, Vectorize).Complete with planning docs, session handoff protocol, and enable scripts for opt-in features.Use when: starting new full-stack project, creating Cloudflare app, scaffolding web app,AI-powered application, chat interface, RAG application, need complete starter, avoid setup time,production-ready template, full-stack boilerplate, React Cloudflare starter.Prevents: service configuration errors, binding setup mistakes, frontend-backend connection issues,CORS errors, auth integration problems, AI SDK setup confusion, missing planning docs,incomplete project structure, hours of initial setup.Keywords: cloudflare scaffold, full-stack starter, react cloudflare, hono template, production boilerplate,AI SDK integration, workers AI, complete starter project, D1 KV R2 setup, web app template,chat application scaffold, RAG st

Meta-skill for selecting correct WorldCrafter skill. Use when unclear which skill to use, request maps to multiple skills, need to orchestrate skills, user asks for guidance, or ambiguous requests need clarification. Provides decision trees and trigger matching for feature-builder (complete features with forms), database-setup (tables/migrations/RLS), test-generator (add tests), route-creator (simple pages/APIs), auth-guard (authentication/authorization), ai-assistant (AI generation/suggestions/consistency), visualization (maps/timelines/graphs), and chatgpt-setup (MCP integration). Do NOT use when request clearly maps to one skill or trigger phrases clearly indicate specific skill.

Implement comprehensive audit logging for all admin actions, capturing user ID, action type, entity changes, IP address, and user agent. Use when tracking system activities or adding audit trails.

U-llm-sdk and claude-only-sdk patterns. Use when working on projects with LLM service, designing LLM integrations, or implementing AI-powered features.

Spawn and manage Claude sub-agents for parallel or delegated tasks. WHEN: User needs parallel operations, multi-step autonomous research, or wants to delegate complex tasks to sub-agents. Use when task can be parallelized or requires independent exploration. WHEN NOT: Simple single-step operations, sequential dependencies, specific file reads (use fs_read_file), writing code directly (do it yourself).