Key Terminology

AI tooling has a lot of jargon. Master these and documentation becomes easy.

• LLM — large language model (Claude, GPT, Gemini, Llama, Hermes…).
• Prompt — the text you send to a model.
• Context window — how much the model can see at once.
• Agent — a program that uses an LLM to choose actions and execute them in a loop.
• Tool / function calling — letting the model invoke code (edit a file, query a DB, browse the web).
• MCP — Model Context Protocol, an open standard for connecting tools and data to AI apps.
• Token — a chunk of text the model processes; models bill by tokens.
• Inference — running the model to produce output.
• Model — a specific trained network (e.g. claude-3-5-sonnet).
• Provider / API — a service that runs models for you (Anthropic, OpenAI, Google, Groq).
• Local model — a model running on your own machine (via Ollama, LM Studio).
• Hallucination — confident but wrong output.
• RAG — retrieval-augmented generation: searching documents and injecting them into context.
• System prompt — fixed instructions that steer the model for the whole session.
• Sandbox — an isolated environment where an agent can run untrusted commands safely.

Example

Prompt:        "Fix the null-check bug in utils.js"
System prompt: "You are a careful senior engineer. Prefer small diffs."
Context:       [utils.js contents, failing test output]
Tool calls:    read_file(utils.js) → edit_file(...) → run_tests()

That one-liner prompt, plus context and tools, is the entire recipe behind a coding agent.