Skill UI

This comprehensive guide outlines the principles, architecture, and best practices for developing specialized, modular skills that extend large language model capabilities. Learn how to integrate domain expertise, multi-step workflows, API tool integrations, and structured resources while managing the context window efficiently.

This skill covers the architecture of intelligent agent memory, including short-term context windows and long-term vector stores. It explores cognitive frameworks like CoALA to organize semantic, episodic, and procedural memory. Focuses on retrieval strategies, chunking, and embeddings to ensure agents remember and recall information effectively across sessions.

This skill optimizes multi-agent systems through coordinated profiling, workload distribution, and cost-aware orchestration. It helps improve agent performance, throughput, and reliability by identifying bottlenecks, managing context windows, and reducing latency. Suitable for complex workflows requiring intelligent coordination and efficiency tracking across multiple AI agents.

Main orchestrator agent that routes tasks across a specialized squad (analyst, architect, builder, reviewer, QA, DevOps). It keeps compressed summaries, stores full artifacts by reference, tracks project state, and conserves its context window.

This skill covers splitting extracted text into chunks suitable for LLM context windows or RAG ingestion. It includes configuring chunk size, overlap, and selecting chunker types like markdown, yaml, or semantic. Users can use the standalone chunk command or enable inline chunking during extraction. It supports token-based sizing for specific models and provides configuration options via CLI, config files, or Python API.

This skill covers strategies for compressing conversation history in AI agent sessions to manage context window limits. It focuses on optimizing tokens-per-task rather than just tokens-per-request. Techniques include anchored iterative summarization, opaque compression, and regenerative summaries. It also addresses artifact trail integrity and evaluation methods like probe-based testing to ensure critical information like file paths and decisions are preserved during compression.

This skill covers the core concepts of context engineering for language models, including system prompts, tool definitions, retrieved documents, message history, and tool outputs. It explains attention mechanisms, context windows, and strategies like progressive disclosure to optimize token usage and agent performance.

A context preservation skill that creates snapshots and verifies data integrity before automatic compression. Protects critical project information from being lost during Claude Code's context window compression, ensuring zero data loss for complex technical projects with multiple skills and dependencies.

This skill specializes in advanced AI context engineering, managing dynamic information flows, vector databases, and knowledge graphs. It assists in building intelligent memory systems, optimizing context windows, and orchestrating multi-agent workflows. Ideal for enterprise-scale AI applications requiring coherent state management, RAG implementation, and semantic retrieval strategies to enhance system performance and reliability.

Context optimization extends effective capacity of limited context windows via compression, masking, caching, and partitioning. It aims to better utilize available capacity rather than increasing window size. Techniques include summarizing context, masking verbose outputs, reusing KV-cache, and partitioning work across sub-agents to reduce costs and latency in AI agent systems.

This skill covers strategies for optimizing Large Language Model context windows to prevent context rot. It includes techniques like summarization, trimming, routing, and prioritization to manage token budgets effectively. Developers can use these methods to build multi-turn conversation systems where context length is a constraint, leveraging tools like tiktoken and LangChain for implementation.

A protocol that provides LLM agents with persistent structural memory of a codebase. It maps the project as a directed graph, tracking not only dependencies (what imports what) but also the 'reasons' (why it connects). This allows agents to understand context, perform impact analysis, and navigate large repositories without loading the entire source code into the context window.