A detailed field guide published on Towards Data Science explores hybrid local-cloud large language model (LLM) workflows, offering a framework for developers to integrate local model privacy with cloud model reasoning. The guide introduces a three-axis coordinate system to map hybrid patterns: local-first versus cloud-first processing, conditional versus always-on cloud triggering, and primary motivation (privacy, cost, latency, or reliability). A case study demonstrates a three-step workflow using Gemma 4 (local) and GPT-5.4 (cloud) for a smart-home scheduling problem. The local model, served with Ollama, processes sensitive private context and converts it into an abstract, anonymised problem, which is sent to the cloud model for complex reasoning without direct access to sensitive data. The local model then translates the cloud model's anonymous results back into user-friendly language. The guide highlights that structured output for smaller local LLMs can incur a 'constraint tax' on task correctness, where smaller models struggle with task correctness despite schema validity, suggesting a need for careful implementation and potential validation and retry logic in production systems.