Ai Integration Skills

Injects retrieved context into prompts correctly — covers context formatting, source attribution, context window budget management, and relevance filtering. Use when building RAG pipelines, document Q&A, search-augmented generation, or any artifact where retrieved chunks must be fed into a Claude prompt reliably.

Implements chunking strategies for documents before embedding or retrieval — covers fixed-size, semantic, recursive, and sliding window approaches with overlap and metadata preservation. Use when building RAG pipelines, document ingestion, vector store population, or any artifact where documents must be split before embedding.

Compress prompts to reduce token count without degrading output quality. Use this skill when the user wants to shorten a system prompt, user prompt, or instruction block; when they say a prompt is "too long", "hitting token limits", or "too expensive to run"; when they want to audit a prompt for bloat; or when they want to make a prompt leaner before scaling it to production. Trigger even if the user says "clean up my prompt" or "tighten this up" — compression is almost always part of what they mean.

Evaluates RAG pipeline quality end-to-end. Use when measuring retrieval precision/recall, answer faithfulness, hallucination rate, or comparing RAG configurations. Trigger for any task involving RAG quality assessment, LLM-as-judge evaluation, RAGAS metrics, groundedness checks, or A/B comparison of retrieval or generation changes.

Parses structured output from LLM responses. Use when extracting JSON from markdown-fenced LLM output, handling partial or streamed JSON, recovering from malformed model responses, or building robust pipelines that consume Claude/OpenAI structured outputs. Trigger for any task involving LLM output parsing, JSON extraction from model responses, stream parsing, or structured output error recovery.

Implements optimized retrieval strategies for RAG and search pipelines. Use when building or improving document retrieval, choosing between similarity search, hybrid search, reranking, or MMR, diagnosing retrieval quality issues, or tuning recall/precision tradeoffs in LLM-powered search. Trigger for any task involving vector search, embedding retrieval, semantic search, BM25, reranking, or RAG pipeline optimization.

Implements the agentic tool execution loop for Anthropic API artifacts — covers tool call detection, dispatch to handler, result injection, and multi-turn continuation until the model signals completion. Use this skill when building any AI-powered artifact that needs Claude to call tools (web search, MCP, custom functions) and loop until a final text response is produced.