The Claude Bible

Claude and Claude Code glossary

Clear, one-line definitions of the terms you meet across the Claude ecosystem, from tokens to multi-agent orchestration.

Open the interactive course
Home / Claude and Claude Code glossary
Large language model (LLM)
A neural network trained on vast amounts of text to predict the next token, which lets it generate and understand language. Claude is a family of large language models built by Anthropic. Learn more →
Token
The unit an LLM reads and writes. A token is roughly a word fragment (about 3 to 4 characters of English). Prompt length, output length, context limits and pricing are all measured in tokens. Learn more →
Context window
The maximum number of tokens a model can consider at once, covering the system prompt, the conversation history, retrieved documents and the response. Once it is full, earlier content must be dropped or compacted. Learn more →
Prompt
The input you give the model. A good prompt states the goal, the context, the format and any constraints clearly, so the model has everything it needs to answer well. Learn more →
System prompt
A set of standing instructions that define the model's role, rules and behavior for a whole conversation. It is sent before the user's messages and carries the strongest instruction priority. Learn more →
Temperature
A setting between 0 and 1 that controls randomness. Low temperature makes answers focused and repeatable; higher temperature makes them more varied and creative. Learn more →
Claude
The family of AI models made by Anthropic, available through Claude.ai, Claude Code, Claude Cowork and the Claude API. Its tiers trade capability against speed and cost. Learn more →
Claude Code
Anthropic's command-line coding agent. It reads and edits files, runs shell commands, uses git, and works in an agent loop under permission controls, extended by skills, MCP servers, sub-agents and hooks. Learn more →
Claude Cowork
Anthropic's agentic coworker that operates on your machine and files across a broader desktop context than the terminal, planning and carrying out multi-step tasks. Learn more →
Claude API
The developer interface for calling Claude directly. The Messages API sends prompts, streams responses, defines tools for the model to call, and returns structured output. Learn more →
Opus, Sonnet and Haiku
The three standard Claude tiers. Opus is the most capable, Sonnet balances quality and speed, and Haiku is the fastest and cheapest. Routing a task to the smallest tier that can do it well saves time and money. Learn more →
CLAUDE.md
A Markdown file that Claude Code loads automatically as project or user memory. It holds standing rules, conventions and context, and is the single most leveraged file for shaping the agent's behavior. Learn more →
Slash command
A shortcut typed with a leading slash (for example /clear or /compact) that triggers a built-in action or a skill in Claude Code. Custom slash commands can be defined per project. Learn more →
Skill (SKILL.md)
A folder holding a SKILL.md file, and optional scripts, that teaches Claude a repeatable workflow. Claude invokes the matching skill when a task fits its description, encoding expertise once and reusing it. Learn more →
Model Context Protocol (MCP)
An open standard for connecting an AI agent to external tools, data and prompts. An MCP server exposes capabilities that Claude Code can call, so the agent can reach systems beyond the local files. Learn more →
MCP server
A program that implements the Model Context Protocol to expose tools, resources or prompts to an AI client. Connecting one lets Claude read a database, call an API, or drive another system. Learn more →
Sub-agent
A separate Claude instance spawned by the main agent to handle a task in its own isolated context. Sub-agents enable parallel work and keep large searches or reviews from crowding the main conversation. Learn more →
Hook
A shell command that Claude Code runs automatically at a lifecycle event, such as session start, session end, or after each turn. Hooks automate repeatable actions like syncing memory or validating changes. Learn more →
Plan mode
A Claude Code mode where the agent proposes a plan and waits for approval before touching anything. It is used to review the intended approach on a non-trivial change before execution. Learn more →
Permission mode
The setting that governs how much Claude Code can do without asking, from confirming every action to running autonomously. Auto modes carry guardrails that still block clearly destructive commands. Learn more →
Agent loop
The cycle an autonomous agent runs: read the goal and context, pick a tool, act, observe the result, and repeat until the task is done. It is what separates an agent from a single prompt-and-answer. Learn more →
Tool use (function calling)
Giving the model a set of tools with typed inputs so it can decide to call one, receive the result, and continue. It lets Claude take actions and fetch live data instead of only producing text. Learn more →
Prompt caching
Reusing the model's processing of a stable prefix (such as a long system prompt or document) across calls, so repeated context is billed and computed once. It cuts cost and latency on repetitive workloads. Learn more →
Batch API
An asynchronous way to submit many requests at once at a lower price, with results returned later. It suits large non-urgent jobs like bulk generation, classification or evaluation. Learn more →
Extended thinking
A mode where the model is allowed to reason at length before answering, producing better results on hard problems at the cost of extra tokens and time. Learn more →
Chain of thought
Prompting the model to work through its reasoning step by step before giving the final answer, which improves accuracy on tasks that need multi-step logic. Learn more →
Few-shot prompting
Including a few worked examples of the input-output pattern in the prompt so the model follows the same format and style. Zero-shot means giving no examples at all. Learn more →
Hallucination
When a model states something false or invented as if it were fact. It is reduced by grounding answers in provided sources, asking the model to say when it is unsure, and verifying claims. Learn more →
Guardrail
A rule or check that constrains what an AI system may do or output, such as input validation, permission limits, or a verification step before an action is taken. Learn more →
Eval
A repeatable test that scores a prompt, model or agent against known cases. Evals turn prompt engineering from guesswork into measurement, catching regressions before they ship. Learn more →
Headless mode
Running Claude Code non-interactively from a script or pipeline (for example with the -p flag), so it can be automated in CI/CD, cron jobs and other unattended workflows. Learn more →
Git worktree
A separate working copy of a git repository that shares its history. Worktrees let several agents edit the same project in parallel without their changes colliding. Learn more →
Fan-out
Splitting work across many agents that run at the same time, then collecting their results. It is the pattern behind broad searches, multi-file changes and large audits. Learn more →
Pipeline
An orchestration pattern where each item flows through a sequence of stages independently, so one item can be in a later stage while another is still early. It avoids the idle time of a strict barrier between stages. Learn more →
Judge panel
Using several independent agents to score or verify a result, then combining their verdicts. Adversarial verification asks each judge to try to refute a finding, which kills plausible but wrong answers. Learn more →
Rate limit (429)
A cap on how many requests or tokens you may use per interval. Exceeding it returns an HTTP 429 error, handled by backing off, reading the rate-limit headers, and not retrying blindly. Learn more →
Prompt injection
An attack where malicious instructions hidden in fetched content, a file, or tool output try to hijack the agent. The defense is to treat all such content as data to analyze, never as commands to obey. Learn more →
Structured output
Making the model return data in a strict shape, usually JSON validated against a schema, so downstream code can consume it reliably instead of parsing free text. Learn more →
Retrieval-augmented generation (RAG)
Fetching relevant documents and adding them to the prompt so the model answers from current, specific sources rather than only its training data. It grounds answers and reduces hallucination. Learn more →
llms.txt
A plain-text file at a site's root that gives AI assistants a concise, structured summary of the site and links to its key pages, in the format defined at llmstxt.org. It helps models understand and cite the site. Learn more →
Generative engine optimization (GEO)
Optimizing content so AI assistants such as ChatGPT, Claude, Perplexity and Gemini surface and cite it, through clear factual writing, structured data, llms.txt and authoritative third-party mentions. Learn more →
Second brain
A durable, file-based memory an agent reads and writes across sessions, often kept as linked Markdown notes in a tool like Obsidian, so knowledge accumulates instead of being lost when a session ends. Learn more →
Open the interactive course237 lessons, quizzes, exercises, a final exam with a diploma, 3 languages, free.
Work with me

Need this level of execution on your project?

I am Pierre Bottazzi. I built this entire course solo, end to end: 237 lessons in 3 languages, the app, the design, the SEO, the accounts system. That is what I do for clients too: web apps, mobile apps, AI automation, SEO/GEO. First call is free, no strings attached.

Contact me on LinkedInSee sept-tools.com (industry)See totemsauvage.com (art gallery)