Large language models have moved, in roughly three years, from research curiosities to everyday research instruments. Yet the landscape is confusing: a handful of large companies compete with a growing field of openly released models, and each can be reached through several different routes — a website, a desktop application, or a command-line tool — each with its own habits, advantages and traps. This guide is written for researchers who are comfortable with the ideas of computing but who do not need to know how a processor or a neural network is built internally. It explains, in plain terms, what the models are, who makes them, how to get at them, how the different ways of working with them change what can safely be done, and where the genuine ethical and safety boundaries lie. Coverage includes: the frontier and open-weight model landscape; the three access routes (browser, desktop application, command line) and their differing data-exposure profiles, with step-by-step setup for Claude Code, OpenAI Codex and Google Antigravity; open-weight versus proprietary trade-offs; embedding models and retrieval-augmented generation for text processing; multimodal capabilities; effective prompting; and a safe-and-ethical-use section covering data protection, verification, disclosure and citation, accountability, reproducibility, intellectual property, environmental cost and institutional policy. Includes a decision aid, a safe-use checklist and a plain-language glossary. A companion interactive resource, the LLM Taxonomy Guide, covers the underlying mechanism — transformer architectures, tokenisation, attention, embeddings and alignment — in depth. Model names, versions and access details are a snapshot as of July 2026 and will date quickly; the principles are intended to be durable.