Every term you'll bump into while using Claude — explained in plain English, no engineering background needed. Listed A to Z. Companion pages: AI Models — Who Makes What and Claude Code in VS Code.
A–Z
| Agent / Agentic | An AI set up to pursue a goal across several steps on its own — plan, act, check, repeat — instead of answering one question at a time. "Agentic" just means "acting like an agent." (The HubSpot connector is this idea made real: Claude reading your CRM and, with approval, updating records.) |
| AGI (Artificial General Intelligence) | The hypothetical future AI that matches humans across every kind of thinking. Today's AI isn't that — but it's the term behind most breathless headlines. |
| AI (Artificial Intelligence) | Software that performs tasks we'd normally call "thinking" — understanding language, summarizing, drafting, answering questions. It's not conscious; it's very sophisticated pattern-prediction. |
| Alignment | The work of making AI actually pursue what people intend — helpful, honest, harmless — rather than gaming its instructions. Anthropic's founding focus. |
| Anthropic | The company that builds Claude, focused on making AI that's helpful, honest, and safe. |
| API (Application Programming Interface) | The "behind-the-scenes" way other software talks to Claude — how developers build Claude into your apps and workflows. You won't use it directly, but your tools might. |
| Artifacts | In the Claude app, a side panel where Claude builds something you can keep and edit live — a document, a table, a working web page, a chart. Great for drafts you'll iterate on. |
| Automation | Work that runs without a person doing each step. Our Tuesday class digest is automation; the goal of the marketing apparatus is more of it. |
| Benchmark | A standardized test used to compare models ("scored X on math problems"). Useful signal, heavily marketed — treat vendor scores with a grain of salt. |
| Bias | Skews inherited from training data that can tilt outputs unfairly. The reason humans review anything that affects people — hiring, pricing, credit. |
| Chain of thought | The step-by-step reasoning a model writes out while working through a problem — what you're seeing when Claude "thinks" before answering. |
| Chatbot | Any software you talk to in conversation. Claude is technically one — but the word usually means the limited support-widget kind, so it undersells what's here. |
| Claude | Anthropic's AI assistant — the product you're talking to. You chat with it in plain language to write, analyze, summarize, brainstorm, and more. |
| Claude Code | A version of Claude that works alongside developers in the terminal to read and write software. (It's what built this knowledge base.) |
| CLAUDE.md | A file of standing instructions in a project folder. Claude Code reads it at the start of every session — our house rules, remembered automatically. |
| Claude Desktop / Mobile | The same Claude as standalone apps on your computer or phone, with extras like connectors and voice. |
| Claude.ai | The website/app where you chat with Claude in a browser — the everyday entry point for most people. |
| CLI (Command Line Interface) | Controlling a program by typing commands instead of clicking buttons. It looks technical; it's just a text box. Claude Code lives here. |
| Cloud | Software running on someone else's servers, reached over the internet. Claude's brain runs in the cloud — your computer just talks to it. |
| Compaction | Summarizing a long conversation to free up context space (/compact in Claude Code). Keeps the gist; fine details can drop — another reason to start fresh per task. |
| Compute | Raw processing power for training and running models — the expensive ingredient in AI. "More compute" is why frontier models cost what they do. |
| Connector | A pre-built link between Claude and an app you use (e.g. HubSpot, Google Drive, Gmail). Connect it once and Claude can read — and sometimes update — that data for you. |
| Context | Everything Claude can "see" in the current conversation — your messages, files you've shared, and its own replies. It uses this to stay on-topic. |
| Context window | The size limit of that memory — how much text Claude can hold at once. Big windows let it read long documents. Past the limit, the earliest parts may drop off. |
| Conversation / Thread | One continuous chat. Claude remembers within it, but starts fresh in a new one — so keep related work in the same thread. |
| Cron job | A task scheduled to run automatically at set times — "every Tuesday at 9am." How our weekly class digest fires without anyone touching it. |
| Deep learning | Machine learning done with many-layered neural networks. The approach behind essentially all modern AI, Claude included. |
| Diff | A side-by-side view of exactly what changed in a file — old vs. new, line by line. How you review Claude Code's edits before trusting them. |
| Embedding | Text converted into a list of numbers that captures its meaning, so similar ideas land near each other. The math behind "find documents like this one." |
| Extended thinking (Reasoning) | A mode where Claude works through a hard problem step-by-step before answering. Slower, but better on complex analysis or multi-step logic. |
| Few-shot / Examples | Showing Claude one or two examples of what "good" looks like before asking it to produce more. Often the single fastest quality boost. |
| Fine-tuning | Specialized extra training to make a model better at one narrow job or voice. Powerful but heavy — most teams get 90% of the benefit from good prompts and Projects instead. |
| Frontier model | The newest, most capable class of AI models at any moment — the leading edge. Fable is Anthropic's current frontier tier. |
| Generative AI | The kind of AI that creates new content — text, images, code — rather than just sorting or scoring existing data. Claude is generative AI. |
| Git | A save-point system for a folder of files — every change recorded, every version recoverable. The seatbelt that makes fast AI edits safe, even used purely locally. |
| GitHub | The website where git projects get stored, shared, and backed up. Think "Google Drive for code folders," with history. |
| GPU (Graphics chip) | The type of chip that trains and runs AI models. Originally for video games; now the most fought-over hardware on earth. |
| Grounding | Tying answers to a trusted source (your files, a live search) so they're factual and traceable, rather than generated from memory alone. |
| Guardrails | Built-in limits that keep AI from harmful, off-brand, or unsafe output. Some are Anthropic's; some you set in your instructions. |
| Hallucination | When AI states something false but confident — a made-up stat, fake citation, or wrong "fact." The #1 reason to verify anything important before it leaves the building. |
| Hook | A small automatic rule in Claude Code that fires at set moments — e.g. "after every file edit, run the checker." Automation around the AI. |
| IDE (Integrated Development Environment) | An app for working on a project's files in one window — editor, file tree, terminal together. Ours is VS Code, with Claude Code running inside it. |
| Inference | Running a trained model to get answers — every reply you get is "inference." The counterpart of training (learning) vs. using (inference). |
| Iteration | Treating the first answer as a draft and refining ("shorter," "more formal," "add a stat"). Claude is built for back-and-forth — use it. |
| JSON | A simple structured-text format programs use to pass data around ({"name": "HCS-212"}). You'll see it when Claude talks to other tools. |
| Knowledge cutoff | The date the model's training data stops. It won't natively know events after that — unless it can search the web or you paste the info in. |
| Latency | The wait before (and while) an answer arrives. Bigger brains think slower — one reason to match the model to the job. |
| LLM (Large Language Model) | The engine under Claude. Trained on enormous amounts of text, it predicts the most fitting next words to form useful answers. "Large" = trained on a lot; "language model" = it works in words. |
| Machine learning | Software that improves from data and examples instead of hand-written rules — the umbrella method all of modern AI sits under. |
| Markdown | Plain-text formatting you can read raw: # heading, **bold**, - bullet. The house format of AI tools — CLAUDE.md is written in it. |
| MCP (Model Context Protocol) | The open "universal adapter" standard that lets Claude plug into outside tools and data sources in a consistent way. It's the plumbing behind many connectors — you'll see the name, but rarely touch it. |
| Memory | Notes Claude keeps across sessions — preferences, facts, corrections. Different from context, which empties when the conversation ends. |
| Model | One specific "brain" version of Claude. Anthropic offers a few, trading speed for depth. When a tool says "choose a model," this is what it means. (See the Models guide.) |
| Model family (Opus / Sonnet / Haiku) | Opus = the most capable, for hard or high-stakes work. Sonnet = the balanced everyday workhorse. Haiku = the fastest and lightest, for quick, simple tasks. Newer numbers (e.g. 4.x) = newer, smarter versions. (Fable is a newer top tier above Opus.) |
| Multimodal | Able to work with more than just text. Claude can read images, screenshots, PDFs, and charts you upload, not only typed words. |
| Neural network | The brain-inspired structure inside a model: layers of simple nodes connected by weighted links. The strengths of those links are the parameters. |
| Open source / Open weights | Models whose internals are published for anyone to run or modify (Meta's Llama is the famous one). Claude is not — you use it as a service. |
| Orchestration | Coordinating several AI steps or agents into one pipeline — deciding who does what, in what order. The "conductor" layer above individual tasks. |
| Parameters | The billions of tiny learned "dials" inside a model, set during training and frozen after. More parameters ≈ a bigger brain — and a slower, pricier one. |
| Permissions | Claude Code's approval gate: it asks before editing files or running commands. Skippable for trusted work — that's a deliberate choice, not a default. |
| Plugin | An add-on that extends a tool with a bundle of capabilities — in Claude Code, a package of skills, connectors, and commands installed together. |
| Privacy / Training data | Whether your inputs are used to train future models. Anthropic's business and API products are not trained on your data by default — but always confirm the setting for the specific tool and plan. |
| Projects | A saved workspace in Claude that holds your files, instructions, and chat history for an ongoing effort — so Claude has the background every time, without re-explaining. |
| Prompt | What you type to Claude — your question, instruction, or request. A clearer prompt gets a better answer. |
| Prompt engineering | The skill of phrasing requests well: giving context, examples, and a clear goal. You don't need to be technical — being specific is 90% of it. |
| Prompt injection | A security risk where hidden malicious instructions inside a webpage or email try to hijack the AI. Matters most when an AI reads untrusted content — be cautious connecting agents to your full inbox. |
| RAG (Retrieval-Augmented Generation) | A technique where the AI looks up relevant info from your documents or database first, then answers using it — so replies are grounded in your real content, not just general training. |
| Rate limit | A cap on how much you can use in a window. Hit it and you'll wait or upgrade. Normal, not an error. |
| Refusal | When Claude declines a request because of its safety rules. Usually fixed by rephrasing or adding the legitimate context it's missing. |
| Repository (Repo) | One project folder tracked by git — the unit our work is organized in. "The CMO repo" = the folder holding all marketing docs and tools. |
| Sandbox | A walled-off place to run code safely, where mistakes can't touch real files or systems. Test in the sandbox; ship outside it. |
| Script | A small program that does one job start to finish — "render this image," "send the digest." The building blocks Claude Code writes for us. |
| Session | One working sitting in Claude Code. Its context lives and dies with it — CLAUDE.md and memory are what carry over. |
| Skill | A saved playbook Claude loads for a specific job — our brand rules, our naming, our workflows. Written once, followed the same way by everyone's Claude. |
| Slash command | Shortcuts typed with / in Claude Code — /model switches brains, /clear starts fresh, /help lists the rest. |
| Streaming | The answer appearing word-by-word as it's generated instead of all at once. Cosmetic, but it's why replies "type themselves." |
| Subagent | A helper Claude spawned to handle one subtask and report back — how big jobs get split and worked in parallel. |
| System prompt | Hidden background instructions that set Claude's role and rules before you say anything (e.g. "You are a marketing assistant"). Often pre-set by whatever tool you're in. |
| Temperature | A dial for how predictable vs. adventurous outputs are — low for consistency, higher for brainstorming. Usually preset; good to recognize the word. |
| Terminal | The app where the command line runs (on a Mac, it's literally called Terminal). Open it, type claude, and you're in Claude Code. |
| Token | The unit AI uses to measure text — roughly ¾ of a word. Length limits (and context windows) are counted in tokens. "100K tokens" ≈ 75,000 words. |
| Tool use (Function calling) | Claude's ability to do things, not just talk — run a calculation, search, or call another piece of software to fetch live data or take an action. |
| Training | The one-time process that sets a model's parameters by showing it enormous amounts of data. It happens before you ever meet the model — your chats don't retrain it. |
| Transformer | The 2017 neural-network design that made modern language models possible — the "T" in GPT. Claude is built on the same family of ideas. |
| Vibe coding | Describing what you want in plain language and letting AI write the actual code — how non-developers now ship working tools. Review still required. |
| Vision | The image-reading ability specifically — paste a screenshot of a dashboard or a slide and ask Claude what it shows. |
| Web search | Claude looking things up on the live internet to get past its knowledge cutoff — with sources you can check. |
| Zero-shot | Asking for something with no examples given — just the request. The counterpart of few-shot. Claude is good at it; examples still help. |
The one-paragraph version: Claude is an AI assistant built on a large language model. You give it a prompt; it answers within its context window, measured in tokens. It can read images and files (multimodal), build editable artifacts, remember an effort in Projects, and — via connectors and MCP — act as an agent inside tools like HubSpot. Its weak spot is hallucination, so ground it in your real data and keep a human check on anything that matters.