Glossary

A system that uses an LLM to make decisions and execute actions autonomously. Unlike a chatbot, an agent can use tools, read files, run code and chain multiple steps without human intervention at each one.

Anthropic's API for accessing Claude models (Haiku, Sonnet, Opus) from any application. Used to build chatbots, agents, text processing pipelines, code analysis and any task requiring an LLM.

Anthropic's official CLI for using Claude directly in the terminal as a development agent. It has filesystem access, can run commands and acts autonomously to implement features, debug code or run tests.

The maximum text (measured in tokens) a model can process in a single call. Claude Sonnet supports up to 200K tokens of context. When the context fills up, the model can't "see" older information from the conversation.

Configuration file that Claude Code reads automatically on startup in a project. Contains instructions, code conventions, project context and rules the agent must follow. Equivalent to persistent agent memory between sessions.

Prompting technique where input/output examples are included before the actual task. With 2-3 well-chosen examples, the model replicates the pattern with high fidelity. Especially useful for controlling the exact output format.

When an LLM generates incorrect or invented information with full confidence. Not a bug but a feature of how statistical models work. Mitigated with RAG, grounding in real documents and output verification.

Shell command that runs automatically before or after Claude uses a tool. Lets you add guardrails: log actions, run the linter after each edit, or block modifications to protected files.

Large-scale language model trained on large amounts of text to predict and generate text. Claude, GPT-4 and Gemini are examples. The emergent "intelligence" comes from patterns learned in training data.

Anthropic's open standard protocol that lets Claude connect to external tools and data sources. An MCP server exposes functions Claude can invoke: database queries, API calls, access to services like Slack or Jira.

Claude API feature that stores parts of the prompt in cache for 5 minutes. Cached tokens cost ~90% less on subsequent requests. Very useful for long system prompts, reference documents or long conversations.

Pattern where relevant documents are retrieved from a database and included in the prompt before generation. Lets the model answer with up-to-date or private information without fine-tuning.

Persistent instructions sent to the model before the user message. Defines the assistant's role, behavior, constraints and context. In Claude API it's the system field of the messages.create call.

Basic unit of text processed by an LLM. Approximately 1 token = 0.75 words in English. Claude API is billed per input and output tokens. A 1000-word text equals roughly 1300 tokens.

Claude's ability to invoke external tools during a conversation. A list of available functions is defined and Claude decides when and how to call them. Foundation of modern agentic systems.