LLM 與 Agent

Research-backed prompting techniques for improved AI response quality (+45-115% improvement). Use when optimizing prompts, enhancing agent instructions, or when maximum response quality is critical. Invoked by /ai-eng/optimize command. Includes expert persona, stakes language, step-by-step reasoning, challenge framing, and self-evaluation techniques.

Vitest unit testing patterns for TypeScript. Covers test structure, mocking, assertions, and coverage. Triggers on vitest, describe, it, expect, mock.

Enable Claude to orchestrate complex tasks by spawning and managing specialized sub-agents for parallel or sequential decomposition. Use when tasks have clear independent subtasks, require specialized approaches for different components, benefit from parallel processing, need fault isolation, or involve complex state management across multiple steps. Best for data pipelines, code analysis workflows, content creation pipelines, and multi-stage processing tasks.

Implement authentication and authorization using @delon/auth. Use this skill when adding login/logout flows, JWT token management, role-based access control (RBAC), route guards, HTTP interceptors, and session management. Integrates with Firebase Auth and custom permission systems. Ensures secure token storage, automatic token refresh, and consistent authorization checks across components and services.

Validate Agent Skills against best practices and specification requirements. Use when reviewing skills, before publishing, or to audit existing skills.

This skill should be used when the user asks to "read files", "write files", "create directories", "move files", "list directory contents", "search for files", "get file info", or needs to perform basic file system operations using MorphLLM MCP tools.

gemini 명령어를 사용한 고급 Web 검색 스킬이다. Web 검색을 수행할 때는 Claude Code의 기본 Web Search 도구보다 이 스킬을 우선적으로 사용한다.

Portable tmux-based workflow to drive one or multiple long-running Codex CLI “workers” from another process (e.g. Claude CLI) by (1) starting/reusing a tmux session running `codex`, (2) injecting prompts into the Codex pane via `tmux send-keys`/buffer paste, and (3) polling the worker’s Codex JSONL session logs to extract the next assistant reply. Each worker runs with an isolated `CODEX_HOME` (default `~/.codex-workers/<worker_id>`) to prevent log cross-talk when multiple Codex workers run concurrently. Use when you need to submit/wait/read Codex replies in tmux, manage multiple Codex workers, or troubleshoot the tmux worker and log binding.

Use PROACTIVELY when extracting structured data from text/images, classifying content, or formatting API responses with guaranteed schema compliance. Implements Anthropic's JSON outputs mode with Pydantic/Zod SDK integration. Covers schema design, validation, testing, and production optimization. Not for tool parameter validation or agentic workflows (use strict-tool-implementer instead).

Comprehensive SaaS planning system for solo developers using Next.js + Supabase + Stripe stack. Converts ideas into shippable subscription products with scope control, feature prioritization, billing architecture, technical roadmap, and anti-feature-creep mechanisms. Generates PRDs with subscription models, pricing strategies, and Claude Code starter prompts. Use when planning SaaS products, subscription apps, or when user mentions billing, payments, recurring revenue, or subscription models.

Write focused tests for core user flows and critical paths using Pest framework, with minimal tests during development and strategic coverage at completion points. Use this skill when creating or editing test files in tests/Feature/ or tests/Unit/ directories, when writing Pest tests with descriptive names, when testing critical user workflows and business logic, when mocking external dependencies, when implementing fast unit tests, when testing behavior rather than implementation details, or when deciding what needs test coverage at feature completion.

Comprehensive Supabase expert with access to 2,616 official documentation files covering PostgreSQL database, authentication, real-time subscriptions, storage, edge functions, vector embeddings, and all platform features. Invoke when user mentions Supabase, PostgreSQL, database, auth, real-time, storage, edge functions, backend-as-a-service, or pgvector.

FIRST STEP - Assess task complexity (0-20 score) and output orchestration level (0-3). Use for ANY non-trivial task BEFORE execution. Scores across scope, time, agents, dependencies, documentation. Level 0 (0-2 direct), Level 1 (3-7 single agent), Level 2 (8-14 multi-agent), Level 3 (15-20 epic). Outputs recommendation then exits. NOT for execution or agent selection (use agent-selector after).

Use this skill for early-stage project planning. It leverages the Gemini CLI to generate high-quality first drafts of Product Briefs, Research Documents, and full PRDs, guiding users from idea to validated requirements. Triggers include "create PRD", "product brief", or "validate requirements".

Document citations and RAG (Retrieval-Augmented Generation) patterns for Claude. Activate for source attribution, document grounding, citation extraction, and contextual retrieval.

Define and run tool-calling patterns for LLMs (schema design, call loops, validation, parallel calls). Use when building function/tool calling workflows or debugging tool selection and arguments; triggers: tool-calling, function-calling, tool schema, tool declaration, parallel function calling.

Complete knowledge domain for Cloudflare Vectorize - globally distributed vector database for buildingsemantic search, RAG (Retrieval Augmented Generation), and AI-powered applications.Use when: creating vector indexes, inserting embeddings, querying vectors, implementing semantic search,building RAG systems, configuring metadata filtering, working with Workers AI embeddings, integratingwith OpenAI embeddings, or encountering metadata index timing errors, dimension mismatches, filtersyntax issues, or insert vs upsert confusion.Keywords: vectorize, vector database, vector index, vector search, similarity search, semantic search,nearest neighbor, knn search, ann search, RAG, retrieval augmented generation, chat with data,document search, semantic Q&A, context retrieval, bge-base, @cf/baai/bge-base-en-v1.5,text-embedding-3-small, text-embedding-3-large, Workers AI embeddings, openai embeddings,insert vectors, upsert vectors, query vectors, delete vectors, metadata filtering, namespace filtering,topK

Design, implement, and debug a custom ChatKit backend in Python that powers the ChatKit UI without Agent Builder, using the OpenAI Agents SDK (and optionally Gemini via an OpenAI-compatible endpoint). Use this Skill whenever the user wants to run ChatKit on their own backend, connect it to agents, or integrate ChatKit with a Python web framework (FastAPI, Django, etc.).

Execute AI security CTF challenges across any competition platform with adaptable workflows for indirect prompt injection, jailbreaks, agent exploitation, and evidence collection with research-grounded techniques

Manager session that launches worker Claude instances in parallel git worktrees for concurrent implementations. Observer subagents monitor workers and report results. Manager evaluates implementations and accepts the best one into the target branch.