Skill UI

A protocol enabling AI agents from different vendors to collaborate on long-running projects, route work to best-suited agents, and improve over time. Uses markdown files with no central server, featuring capability cards, contracts, chronicles, and self-amendment mechanisms for cross-vendor optimization and constitutional evolution.

This skill covers principles for identifying tasks suited to LLM processing, designing effective project architectures, and iterating rapidly using agent-assisted development. It applies to batch pipelines, multi-agent systems, and interactive applications. Key concepts include task-model fit recognition, manual prototyping, pipeline architecture, file system state management, and structured output design for reliable parsing.

Caching strategies for LLM prompts, including Anthropic prompt caching, response caching, cache invalidation, and Cache Augmented Generation (CAG) to reduce latency and cost.

Master advanced prompt engineering techniques to maximize LLM performance, reliability, and controllability. This skill covers few-shot learning, chain-of-thought prompting, prompt optimization, template systems, and system prompt design. It helps in designing complex prompts for production applications, optimizing consistency, and debugging inconsistent outputs. Ideal for developers building AI assistants and production-grade LLM applications.

Protect AI systems from prompt injection attacks by implementing input sanitization, tool permissioning, and output validation. Essential for securing LLM applications, RAG pipelines, and agentic workflows against indirect and direct attacks.

A governance skill designed to secure AI agent tool calls using advanced authorization policies. It implements Cedar policies to control which tools agents can access and under what conditions. Key features include running in 'Shadow mode' for safe observation, 'Enforce mode' for strict blocking, and generating cryptographically signed Ed25519 receipts to provide an immutable, tamper-evident audit trail for every policy decision.

PydanticAI is a robust Python framework that integrates Pydantic's type-checking and validation capabilities into Large Language Models (LLMs). It allows developers to build production-ready AI agents that guarantee structured, validated outputs, moving beyond simple raw strings. Key features include supported tool calling, dependency injection for testability, and multi-model compatibility across major providers like OpenAI, Anthropic, and Gemini. Use this when reliability and correct data structure are paramount.

Delegate a bounded coding task to the Qoder CLI implementer, then review its diff, re-run project gates, and commit the verified result yourself as orchestrator.

This comprehensive workflow guides the end-to-end implementation of Retrieval-Augmented Generation (RAG) systems. It covers critical stages, including selecting optimal embedding models, setting up vector databases, designing chunking strategies, implementing hybrid retrieval, integrating Large Language Models (LLMs), and performing rigorous evaluation. Use this when building knowledge-grounded AI applications or advanced semantic search systems.

Build and operate Retrieval-Augmented Generation (RAG) systems using vector stores, embedding pipelines, and hybrid search. This skill covers ingesting documents, generating embeddings, storing in vector databases, and serving grounded LLM responses. It includes setting up embedding pipelines, vector store infrastructure, and deploying hybrid search (dense + sparse) to reduce hallucinations and improve semantic search over large document collections.

Monitor and evaluate Retrieval-Augmented Generation (RAG) systems using measurable metrics. This skill covers retrieval quality checks like Recall@k, hallucination detection, and groundedness scoring. It enables continuous regression testing and production monitoring for AI pipelines using Python libraries like RAGAS and OpenTelemetry.

Execute exactly one explicitly assigned YYLO Ledger task through the Ralph loop to a validated, queued commit using the yy CLI. Handles task admission, preflight and finish while never pushing, deploying, merging or releasing.