Claude Skills

Use when validating golden dataset quality. Runs schema checks, duplicate detection, and coverage analysis to ensure dataset integrity for AI evaluation.

[CONFIG] Configure OrchestKit settings. Use when customizing MCP servers, plugin options, or preferences.

Use when conversation context is too long, hitting token limits, or responses are degrading. Compresses history while preserving critical information using anchored summarization and probe-based validation.

Multi-directory context patterns for monorepos. Use when working with --add-dir, per-service CLAUDE.md, or separating root vs service context

5 Whys, Fishbone diagrams, Fault Tree Analysis, and systematic debugging approaches. Use when investigating bugs, analyzing incidents, or identifying root causes of problems.

CLIP, SigLIP 2, Voyage multimodal-3 patterns for image+text retrieval, cross-modal search, and multimodal document chunking. Use when building RAG with images, implementing visual search, or hybrid retrieval.

Production hybrid search combining PGVector HNSW with BM25 using Reciprocal Rank Fusion. Use when implementing hybrid search, semantic + keyword retrieval, vector search optimization, metadata filtering, or choosing between HNSW and IVFFlat indexes.

Query decomposition for multi-concept retrieval. Use when handling complex queries spanning multiple topics, implementing multi-hop retrieval, or improving coverage for compound questions.

LangGraph Functional API with @entrypoint and @task decorators. Use when building workflows with the modern LangGraph pattern, enabling parallel execution, persistence, and human-in-the-loop.

LangGraph parallel execution patterns. Use when implementing fan-out/fan-in workflows, map-reduce over tasks, or running independent agents concurrently.

Multi-agent coordination and synthesis patterns. Use when orchestrating multiple specialized agents, implementing fan-out/fan-in workflows, or synthesizing outputs from parallel agents.

Temporal.io workflow orchestration for durable, fault-tolerant distributed applications. Use when implementing long-running workflows, saga patterns, microservice orchestration, or systems requiring exactly-once execution guarantees.

LLM fine-tuning with LoRA, QLoRA, DPO alignment, and synthetic data generation. Efficient training, preference learning, data creation. Use when customizing models for specific domains.

LLM function calling and tool use patterns. Use when enabling LLMs to call external tools, defining tool schemas, implementing tool execution loops, or getting structured output from LLMs.

High-performance LLM inference with vLLM, quantization (AWQ, GPTQ, FP8), speculative decoding, and edge deployment. Use when optimizing inference latency, throughput, or memory.

HyDE (Hypothetical Document Embeddings) for improved semantic retrieval. Use when queries don't match document vocabulary, retrieval quality is poor, or implementing advanced RAG patterns.

Statistical and quality drift detection for LLM applications. Use when monitoring model quality degradation, input distribution shifts, or output pattern changes over time.

PII detection and masking for LLM observability. Use when logging prompts/responses, tracing with Langfuse, or protecting sensitive data in production LLM pipelines.

ROI, NPV, IRR, payback period, and total cost of ownership analysis for investment decisions. Use when building financial justification for projects, evaluating SaaS investments, or comparing alternatives.