Updated July 2026. The tools move faster than the words, so treat definitions as current-as-of, not carved in stone. The concepts underneath change slowly. That is why they are worth learning.
This is the plain-language version, in the voice of this book, of the terms you will hear thrown around, not an encyclopedia. Where I have an opinion, I have kept it.
Model
The actual AI underneath everything, the thing that was trained and does the reasoning. Claude is a model, or really a family of them. When people compare "which model is better," this is what they mean. For most of your work the model is not the bottleneck. The context you give it is.
Prompt
What you type. The instruction or question you send the AI. That is all a prompt is, despite the mystique. A good prompt is clear about what you want. It is not a magic spell, and no phrase unlocks a secret better AI.
System prompt
A set of instructions that sits above the conversation and shapes how the AI behaves throughout it, without you retyping it each turn. Think of it as the standing instructions for a session rather than the individual thing you ask. Your CLAUDE.md is a close cousin of this idea, made persistent and yours.
Context window
How much the AI can hold in its attention at once, counted in tokens. Everything it is currently working with, your instructions, the conversation, any files it is reading, has to fit inside this window. Big now, but not infinite, which is why pointing an AI at the right focused material beats dumping everything at it.
Token
The unit the AI reads and counts in. Roughly a word or a chunk of one. You do not need to think about tokens day to day. You only need to know they are what the context window is measured in, so "too much at once" is a real limit, not a vibe.
Prompt engineering
The craft of wording your prompt well. Real, but oversold. It was the whole game in the early days and it is still the expensive way to work if your work repeats, because a clever prompt lives in one conversation and dies with it. You rebuild it from scratch every time. Fine for one-offs. A trap for anything you do more than twice.
Context engineering
The actual leverage. Instead of re-explaining yourself in every conversation, you write down who you are and how you work in a place the AI reads automatically before every task. You build the world it works inside, once. This is what most of this book is about, and it is the difference between an assistant that forgets you overnight and one that shows up already knowing you.
Loop engineering
The next step past context. The AI stops running once and waiting for you, and instead drafts, checks its own work against criteria you set, revises, and only then brings it to you, or runs on a schedule without being asked at all. Your job shifts from doing the work to defining the standard the work has to pass. Covered in the bonus chapter.
Second brain
A term from the note-taking world. It means a system outside your head that holds what you know so you are not carrying it all in working memory. Context engineering is the same instinct pointed at your AI: a second brain the machine reads before every task, not one you reread yourself. Where it lives does not matter much, only two things do: the AI can read it (local files, a connector, or MCP) and it has a structure the AI can navigate (named files or pages, one topic each, not one giant dump). A folder of markdown files, an Obsidian vault, a GitHub repo, Notion, or Google Drive all qualify. Use the one you already use rather than migrating for tidiness. It is also what makes changing tools cheap, because when a surface breaks or you switch, your context survives in the shared place and only the tooling has to be rebuilt.
Surface
My word, in this book, for the three forms Claude comes in: the Chat, Cowork, and Code. They are not the same tool with different skins. Each one can see and remember a different amount of your actual work. Choosing the right surface for a job matters more than choosing the right words.
Chat
The browser or mobile box you already know. Fast, everywhere, and built to forget. It can reach your connected external tools, but it cannot see the files on your own computer. Perfect for quick questions, throwaway drafts, and capturing ideas on the move. Wrong for anything that needs to open your local files.
Cowork
The desktop app. Because it lives on your computer, it can see your files and folders and act on them, and it reaches your connected external tools too. The surface for work that touches documents and data you already have on your machine.
Code
The surface built for developers but useful to far more people than that. Like Cowork it sees your local files, and like all three it reaches your external tools. Its real difference is that it lives inside your project folder and its history, holds your always-on CLAUDE.md and your agent teams, and can run real work there rather than only describe it. The name scares people off. Underneath it is the surface that moves into your project and works there.
CLAUDE.md
A plain text file, named exactly CLAUDE.md, that the AI reads before every task on that surface. Your always-on operating brief. Starting out, put everything in it: who you are, your customers, your voice, your rules. As your setup grows it becomes a router rather than a warehouse: the always-read part holds who you are, your house rules, and a map of where deeper context lives, while the deep context moves into scoped files and connected sources fetched when relevant. The reason is that the AI reads this on every task, so it should carry what every task needs and point to the rest. The single highest-leverage file in your whole setup, and one you maintain, not one you build and forget.
Project instructions
Scoped context. Where CLAUDE.md is who you are in general, project instructions are the context for one specific kind of work, attached to that work so it does not bleed into everything else. Your proposal context stays with proposals, your content context stays with content.
Project
A workspace, available in the Chat and in Cowork, where you attach standing instructions and upload files. Every conversation inside the project automatically reads that project knowledge. The catch nearly everyone hits: separate conversations in the same project do not see each other's transcripts. A project shares what you put in it, the instructions and files, not what you say in it, the chats. If something from a conversation matters later, move it into the project knowledge yourself.
Memory (the types)
There is no single "memory." There are a few kinds and they behave differently. Conversation memory is alive while a chat is open and gone when it closes. The memory feature distills your conversations into a summary profile that gets carried into future chats, which is real but is a summary, not full recall. Projects share the instructions and files you put in them, though not the transcripts of the chats inside them. CLAUDE.md and project instructions are context you write on purpose and that gets read exactly. The whole point of this book is moving your important context out of the kinds that vanish or only summarize, and into the kinds you build and control.
Memory feature
The layer that summarizes your past conversations, roughly once a day, into a profile Claude carries into future chats. It is why Claude can feel like it "sort of remembers" you. Keep the word summary in mind: it may carry forward that you are a consultant with a certain tone, but not the exact thing you decided three days ago. Useful, not a substitute for context you deliberately write down.
Conversation memory
The vanishing kind. What the AI holds during a single chat. You can reopen that same chat and it is still there, but none of it carries into the next one you start. This is the memory behind the phrase "my AI keeps forgetting me." It is not broken. It was built to be temporary. The fix is to stop relying on it for anything you want to keep.
Auto-memory
The notebook the AI writes about you, on its own, as it learns things across your work together. You do not author it, it accumulates. It is why a well-used setup gets better the longer you use it, remembering decisions and standing facts so you stop re-establishing them.
Skill
A written-down procedure plus your house rules for doing a specific job your way. Where context is who you are, a skill is how you do one kind of work. The tell that you need one: you keep giving the AI the same corrections every time. Those repeated corrections are an unwritten skill.
Agent
A Claude instance with a role, its own context, and a job it can run without you steering every step. Not magic, not autonomous in the movie sense, and only as good as the context and skills you gave it. The difference between a chat you drive turn by turn and a teammate you hand a job to.
Subagent
An agent you hand a big, self-contained job to so it runs on its own, in its own clean context, and brings back a result. The counterpart to keeping work in your main session. Rule of thumb: big and independent goes to a subagent, small and iterative stays with you.
Handoff
The moment work passes from one agent, or one step, to the next. Agents can absolutely hand off to each other directly, and independent work does not need you in between at all, just review it at the end. The advice in this book is narrower: on dependent chains, where one agent's output becomes the next one's input, keep a human at the handoff while you are learning, so an early mistake gets caught where it happens instead of being built on top of and discovered three steps too late. You step out once you have built checks that can catch it for you.
Connector
The everyday name for a plugged-in connection between Claude and one of your outside tools: your Notion, your Drive, your calendar. You click it on, Claude can read and work with that tool directly, and you stop copying things across by hand. Under the hood a connector is MCP at work, but "connector" is the word you will actually meet in the product.
MCP
Model Context Protocol. A standard way of connecting the AI to outside systems, your Notion, your Drive, your other tools, so it can read and work with them directly instead of you copying things across by hand. In everyday use you meet it as connectors. The thing to know: all three surfaces can use connectors, the Chat included, so reaching your external tools is not what separates them. It is how any surface reaches beyond itself.
RAG
Retrieval-augmented generation. A technical term for giving the AI the right relevant material to read before it answers, pulled from a larger store of your documents, so its answer is grounded in your actual information rather than its general training. You benefit from this idea every time you point the AI at your own files. You rarely need to say the acronym.
Hallucination
When the AI states something confidently that is simply not true. It happens because the model is producing plausible text, not looking up facts, unless you have grounded it in real material. The main defense is context: the more it is working from your actual documents, the less room it has to invent.
Grounding
Tying the AI's answer to real source material, your files, your data, so it is working from facts instead of filling gaps with plausible-sounding guesses. The practical antidote to hallucination, and most of context engineering is grounding by another name.
Fine-tuning
Actually retraining a model on your own data to change its underlying behavior. It sounds like what you want and it almost never is. It is expensive, technical, slow to update, and for the vast majority of operators, context and skills get you everything you thought fine-tuning would, without any of the cost. If someone tells you you need to fine-tune, be skeptical. You probably need better context.
Context rot
Also called drift. When your standing context slowly goes stale, or when a long session accumulates so much clutter that the AI starts losing the thread of what matters. The fix is maintenance: keep your context files current, and start fresh sessions for fresh work rather than dragging an endless conversation behind you. A setup is a garden, not a monument.
Orchestration
Coordinating multiple agents or steps so work flows through them in the right order. This can be automated: agents can run in parallel, and agent teams can coordinate and message each other directly without a human in the middle. Those are real, common patterns. What this book recommends is that while you are learning, you stay the orchestrator on your dependent chains, standing at the handoffs deciding what goes forward, until you have built checks good enough to take over. Doing it yourself first is how you learn what a safe handoff even looks like.
Harness
The machinery around the model that turns it from a text generator into something that can do work: the tools it can call, the files it can touch, the permissions it runs under, the loop that feeds results back to it. Claude Code is a harness, and so is every agent setup you build. Worth knowing because when something misbehaves, the cause is more often the harness, what the model can see and do, than the model itself.
AI OS
The whole thing, assembled. Surfaces, context, skills, and agents wired together so work flows through them with you doing the judgment and the system doing the repetition. Not a product you buy. A wiring of boring parts, which is exactly why you can build your own.