Skill UI

A set of behavioral guidelines derived from Andrej Karpathy's observations to help users writing, reviewing, or refactoring code with Large Language Models (LLMs). It guides developers to avoid overcomplication, make surgical changes, explicitly surface assumptions, and define verifiable success criteria, promoting clean, simple, and robust code development.

This skill guides the development of EarLLM One, a complex Android application. The app captures real-time audio from Bluetooth earbuds, transcribes the voice using STT, sends the text input to a Large Language Model (LLM) via a robust pipeline, and plays the response back using TTS. It requires deep knowledge of Kotlin, Android development, Bluetooth profiles (A2DP/HFP), and advanced audio routing.

An expert guide for developing production-grade AI applications, chatbots, and generative UIs using the Vercel AI SDK, Next.js, and React. Covers core concepts including streaming text (`streamText`), managing chat state (`useChat`), implementing complex tool/function calling, and ensuring robust structured data generation (JSON schemas) across multiple LLM providers.

This blueprint guides the creation of F.R.I.D.A.Y., a tactical, low-latency voice assistant inspired by Iron Man. It utilizes the specialized Pipecat framework to orchestrate a comprehensive, real-time audio pipeline: STT (OpenAI Whisper/GPT-4o) -> LLM (Gemini) -> TTS (OpenAI Nova). Ideal for developers needing to build advanced, multi-provider, conversational voice applications.

This skill provides methods for reliably extracting typed, validated, and structured data from Large Language Models (LLMs), moving beyond simple text parsing. It covers best practices across major APIs (OpenAI's response_format, Anthropic's tool_use, Google's responseSchema) and local models. Key workflows include defining schemas (Pydantic/Zod), implementing constrained decoding, and building robust retry loops for production-grade data pipelines.

This command executes the entire AgentHub lifecycle in a single, chained invocation. It automates complex development tasks by sequentially handling initialization, capturing performance baselines, spawning multiple specialized agents, evaluating the results (via metrics or LLM judging), and finally presenting a winner for merging. Use this when running an end-to-end competition or complex refactoring task.

AgentHub spawns parallel git worktree agents that compete on tasks, evaluates their outputs via metrics or LLM judges, and merges the top branch for optimizations, research, or content variations.

This guide assists developers in designing and architecting Model Context Protocol (MCP) servers that integrate deeply with LLMs like Claude. It is not a coding skill, but a comprehensive consultation process that guides users through use case interrogation, determines the optimal deployment model (e.g., Remote HTTP, MCPB), and selects the appropriate tool-design pattern for seamless, robust integration.

This skill guides developers on building LLM-powered applications using the Claude API, Anthropic SDKs, or the Agent SDK. It provides comprehensive guidance on language detection, best practices (like default model versions and streaming), and choosing the correct development surface (API call, Workflow, or Agent) based on the use case.

Behavioral guidelines for writing, reviewing, and refactoring code with LLMs, emphasizing minimalism, clear assumptions, surgical edits, and verifiable success criteria to avoid common machine-assisted coding mistakes.

Provides a comprehensive reference for migrating deprecated and unsafe class component lifecycle methods (such as componentWillMount, componentWillReceiveProps, and componentWillUpdate) to modern, React 18 compliant patterns. This guide is essential when updating legacy codebases, helping developers correctly identify the semantic purpose of the method and choose the correct replacement hook (e.g., componentDidMount or getDerivedStateFromProps), thus preventing common migration errors.

This skill guides the integration of Arize AX tracing into existing applications, especially complex LLM applications. It follows a structured, two-phase, agent-assisted flow: first analyzing the codebase (read-only) to understand dependencies, and then implementing instrumentation. It supports multi-language environments (Python, TypeScript, Java, Go) and integrates deeply with OpenTelemetry and various LLM providers, ensuring comprehensive observability without altering core business logic.