LLM & Agents

Role definition for code reviewer agents. Defines review criteria, feedback patterns, and approval workflows. Use to understand how to review PRs and provide constructive feedback.

Evaluation framework patterns for RAG and LLMs, including faithfulness metrics, synthetic dataset generation, and LLM-as-a-judge patterns. Triggers: ragas, deepeval, llm-eval, faithfulness, hallucination-check, synthetic-data.

Generate and update HTML dashboards for LLM usage (Claude and Gemini). Use when the user wants to visualize their AI coding assistant usage statistics, view metrics in a web interface, or analyze historical trends.

Wrapper for GitHub CLI (gh) in Claude.ai. Covers installation, authentication, and common operations like pushing files, creating branches, PRs, issues, and releases.

Research file storage conventions and templates for dokhak agents. Use when: (1) saving research results from research-collector or researcher agents, (2) reading cached research files, (3) checking if research exists for a section. Provides directory structure, file format templates, and naming conventions.

Converts all headings in a document into normal paragraphs while keeping the original font size, weight, color, and other formatting intact.

Identifies unreachable functions, unused variables, and abandoned code in binary programs. Use when optimizing binary size, analyzing code coverage, or investigating abandoned functionality.

Phase 2-5 autonomous execution guidance - Activate when Claude needs to decide technical choices during implementation/testing/review/release phases

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.

Best practices for creating, indexing, and managing Bluera Knowledge stores. Covers when to use clone vs crawl vs folder, naming conventions, indexing strategies, storage management, background job monitoring, and handling indexing failures.

This skill should be used when the user asks about "R2", "D1", "KV", "Durable Objects", "Queues", "Vectorize", "Hyperdrive", "Workers Analytics", "Email Routing", "Browser Rendering", or discusses Cloudflare platform services, storage options, database choices, when to use which service, or integration patterns between Workers and platform products.

A simple example skill that demonstrates Claude Code skill structure

Creates PhD-grade expert agents when no suitable agent exists.Produces highly specialized agents with deep domain expertise.

PostgreSQL + pgvector + PGroonga で GraphRAG を構築。自然言語でプロジェクトを説明すると、最適なEntity/Edge型、スキーマ、クエリパターンを設計・生成。Use when the user mentions "GraphRAG", "knowledge graph", or wants to build RAG with relationships.

Use when you need to fine-tune(ファインチューニング) and optimize LangGraph applications based on evaluation criteria. This skill performs iterative prompt optimization for LangGraph nodes without changing the graph structure.

Use when working with DeepResearchAgent/AgentOrchestra - a hierarchical multi-agent system for complex task solving, deep research, and automation. Achieved 83.39% accuracy on GAIA benchmark. Includes planning agents, research agents, browser automation, and MCP tool integration.

Recover from crashed, failed, or interrupted Claude Code sessions. Use this skill when: session crashed during multi-agent parallel execution, need to determine what work was completed vs incomplete, want to generate resumption commands for interrupted tasks, recovering from context window exhaustion, or handling session handoffs. Analyzes agent logs, verifies on-disk state, and creates resumption plans with ready-to-execute Task() commands.

Extract and persist learnings from the current session. This skill should be used when the user wants Claude to analyze the conversation and capture insights, patterns, preferences, or knowledge discovered during work. Triggers on requests like "learn from this session", "what did we learn", or "capture insights".

Automatically generate test scaffolding when user writes new code without tests or mentions needing tests. Supports unit, integration, e2e, and data tests for PHP and JavaScript. Invoke when user mentions "tests", "testing", "coverage", "write tests", or shows new untested code.

Use when completing tasks, implementing major features, or before merging to verify work meets requirements - dispatches core:code-reviewer subagent to review implementation against plan or requirements before proceeding