Every Claude Concept, explained
All 75 terms from my video, from tokens and training to Claude Code and safety, each defined in one line. Search them, jump to a chapter, or grab the one-page cheat sheet.
01The engine18 terms
- 01
Artificial intelligence
Umbrella term for software that does things we used to think needed a human brain.
- 02
Machine learning
AI that learns patterns from millions of examples instead of hand-written rules.
- 03
Neural network
Layers of tiny math units passing numbers; learning adjusts the strength of each connection.
- 04
Large language model
A neural network trained on huge amounts of text to read and write language. Claude is one.
- 05
Pretraining
First training stage: the model guesses the next piece of text, trillions of times.
- 06
Parameters
The connection strengths (weights). A model's knowledge is spread across billions of them.
- 07
Tokens
Chunks of text a model reads and writes, about three quarters of a word each.
- 08
Next-token prediction
Claude writes one token at a time: score every option, pick one, repeat.
- 09
Context window
Everything Claude can see at once. Its working memory, big but finite.
- 10
Knowledge cutoff
The date after which a model hasn't seen any data.
- 11
Hallucination
A fluent, confident answer that's wrong. Check numbers, quotes and citations.
- 12
Temperature and sampling
How the next token is picked. Low temperature is focused, high is more varied.
- 13
Post-training
Second training stage that turns an autocomplete into an assistant that follows instructions.
- 14
RLHF
Reinforcement learning from human feedback: people rate answers, the model moves toward the preferred ones.
- 15
Constitutional AI
Anthropic's method: Claude learns to critique and improve its answers against a written set of principles.
- 16
Anthropic
The AI safety and research company that makes Claude.
- 17
Model family
Haiku (fast), Sonnet (balanced), Opus (harder problems), Fable (most demanding). Bigger costs more.
Deep dive: Claude Fable 5 for UI Design: Beautiful Output Every Time - 18
Multimodal
Claude can take in images, screenshots, charts, handwriting and PDFs, not just text.
02Talking to Claude12 terms
- 19
Prompt
What you send Claude. Better input, better output.
- 20
System prompt
Hidden instructions from whoever built the app: who Claude is, its tools, its rules.
- 21
Prompt engineering
Writing prompts that work: give context, the audience, an example of good, and what to avoid.
Deep dive: How to Prompt AI for Better UI Design - 22
Few-shot examples
Show a couple of finished examples and Claude matches their format and tone.
- 23
Extended thinking
Claude reasons step by step before answering, which helps with math, code and planning.
- 24
Effort
How hard Claude works on a task. Low is quick and cheap, high digs deeper.
Deep dive: Claude Fable 5 for UI Design: Beautiful Output Every Time - 25
Conversations are stateless
Claude re-reads the whole conversation every turn; it doesn't remember between messages.
- 26
Projects
A home base of files and instructions that every chat inside it can see.
- 27
Memory
Preferences and context Claude carries across conversations. You can view, edit or turn it off.
- 28
Styles
Change how Claude writes (concise, formal, your own voice) without changing what it knows.
- 29
Web search
Claude looks things up live and answers with citations, getting around the knowledge cutoff.
- 30
Research
Web search turned up: many searches over several minutes, cross-checked into a cited report.
03Claude makes things5 terms
- 31
Artifacts
Apps, dashboards, diagrams and documents Claude builds live in a panel next to the chat.
- 32
Publishing and sharing
Artifacts are private by default; publish one to share a link, or keep it in your team.
- 33
Artifact capabilities
Artifacts that read live data, save shared state, know the viewer, or ask Claude questions.
- 34
Files and documents
Real spreadsheets, slide decks, Word documents and PDFs you can open in Office.
- 35
Claude Design
A canvas for screens, prototypes and decks built with your design system.
Deep dive: Claude Design Tutorial for Beginners (Full Walkthrough)
04Claude reaches out9 terms
- 36
Tool use
How Claude does things: it requests a tool, the app runs it, the result comes back.
- 37
Agents and the agent loop
Tool use in a loop: decide, act, check the result, repeat until the job is done.
- 38
MCP
Model Context Protocol: an open standard for connecting AI to tools and data. USB-C for AI.
Deep dive: The Best MCP Servers for Designers (2026) - 39
Connectors
One-click MCP: connect email, calendar, docs and more so Claude can read and act in them.
- 40
Skills
Folders of instructions (and scripts) that teach Claude one job; loaded only when needed.
Deep dive: Claude Skills for Designers - 41
Plugins
One install that bundles skills, commands, agents, hooks and connectors.
- 42
Claude in Chrome
An extension that lets Claude navigate, click and fill forms in your own browser.
- 43
Computer use
Claude controls the mouse and keyboard from screenshots, so it can use apps with no API.
- 44
Claude in Slack
Tag Claude in a thread and it reads the context and helps without anyone switching apps.
05Claude Code17 terms
- 45
Claude Code
Claude as an agent on your computer: reads files, edits code, runs commands and tests.
Deep dive: How I Built an iOS App With Claude Code as a Designer (and Shipped It) - 46
Where it runs
Terminal, the desktop app's Code tab, the web, VS Code and JetBrains. Same agent.
- 47
CLAUDE.md
A project file Claude reads every session: how to run, test, your conventions, what not to touch.
Deep dive: How to Write an AGENTS.md File for Design Projects - 48
Permission modes
How much Claude can do without asking, from asking every time to bypass in a sandbox.
- 49
Plan mode
Read-only: Claude explores and writes a plan, and changes nothing until you approve.
- 50
Slash commands
Shortcuts typed with /. Some built in, and you can write your own.
- 51
Hooks
Your scripts that run at set moments, every single time, e.g. format code or block commands.
- 52
Subagents
Helpers with their own fresh context window that do a focused job and report back.
- 53
Sessions and worktrees
Run several sessions at once; worktrees give each its own copy of the repo on its own branch.
- 54
Background and cloud sessions
Sessions that keep working in the cloud while your laptop is closed; check in from your phone.
- 55
Routines and loops
Routines run tasks on a schedule; the loop command reruns a prompt on an interval.
- 56
Compaction
When the context window fills, Claude summarizes older history and keeps going.
- 57
Output styles
How Claude Code narrates its work: concise, explanatory, or a learning mode.
- 58
Mods
Small plugins that change the tool itself: status lines, panels, banners, notifications.
- 59
Workflows
Scripts that coordinate many agents at once for big, thorough jobs.
- 60
Code review
Claude checks a diff for bugs at the depth you choose; ultra runs a deep multi-agent review.
- 61
Fast mode
The same top model with faster output, for when waiting is the bottleneck.
06Building with Claude7 terms
- 62
The API
How developers send messages to Claude from their own code (the Messages API).
- 63
Pricing per token
You pay per token in and out; output and bigger models cost more.
- 64
Prompt caching
Reuses the processed version of a repeated prompt, so requests are faster and much cheaper.
- 65
Batch processing
Send a big pile of requests at once and get them back later at a lower price.
- 66
Agent SDK
The engine behind Claude Code, for developers building their own agents.
- 67
Managed Agents
Anthropic hosts the agent and its sandbox, so you run agents without managing servers.
- 68
Evals
Test cases that measure how well your prompt or agent performs, so you know a change helped.
07Keeping it safe7 terms
- 69
Alignment
Making sure AI does what people intend and stays honest and safe as it gets more capable.
- 70
Prompt injection
Hidden instructions in content Claude reads, like a web page or email, trying to hijack it.
- 71
Jailbreaks
Deliberate attempts to trick a model into ignoring its rules.
- 72
Interpretability
Looking inside a model to understand how it thinks. Neuroscience for AI.
- 73
Responsible Scaling Policy
Anthropic's commitment that stronger safety measures must exist before more capable models.
- 74
Model welfare
Research into whether models like Claude could have experiences that matter.
- 75
Usage policy
The rules for what Claude can and can't be used for.
No concepts match that search.
Common questions
What is a context window in Claude?
The context window is everything Claude can see at once: your messages, its replies, any files you shared and the instructions it was given. It works like working memory. Claude's window holds hundreds of thousands of tokens, enough for whole codebases or several books, but it is finite, and anything outside it isn't there.
What is the difference between MCP, connectors, skills and plugins?
MCP, the Model Context Protocol, is an open standard Anthropic created for connecting AI to tools and data. Connectors are MCP made one-click inside the Claude apps. Skills are folders of instructions, sometimes with scripts or templates, that teach Claude one specific job, and plugins bundle skills, commands, agents, hooks and connectors into a single install.
What are Claude's models: Haiku, Sonnet, Opus and Fable?
Claude is a family of models. Haiku is the small, fast one, Sonnet is the balanced everyday model, Opus is for harder problems, and Fable sits at the top end for the most demanding work. Bigger models are smarter, slower and cost more per token, so picking one is a tradeoff.
What is Claude Code and where does it run?
Claude Code is Claude as a full agent working on your computer. It reads your files, writes and edits code, runs commands and tests, and fixes what breaks. You can use it in the terminal, in the desktop app's Code tab, on the web, and inside editors like VS Code and JetBrains.
What is CLAUDE.md?
CLAUDE.md is a file in your project that Claude reads at the start of every session. It holds what a new teammate would need: how to run and test the app, your conventions, and what never to touch. When you catch yourself correcting Claude about the same thing twice, add a line to it.
What is the difference between hooks and CLAUDE.md instructions?
Instructions in CLAUDE.md are guidance Claude follows most of the time. Hooks are your own scripts that run at set moments, like before Claude edits a file, after it runs a command or when it finishes, and they run every single time. Use hooks for things that must always happen, like auto-formatting code or blocking dangerous commands.
What is prompt injection?
Prompt injection is when text Claude reads, like a web page, an email or a file, contains hidden instructions trying to hijack it. Claude is trained to treat that content as data rather than commands. It is also why agents ask before sensitive actions and why you shouldn't give an agent more access than a task needs.
What is an AI agent?
An agent is tool use in a loop. It is given a goal, decides what to do, uses a tool, looks at what happened and decides the next step, over and over until the job is done. Ask a chatbot to fix a bug and it suggests a fix. Ask an agent and it changes the code, runs the tests and fixes what fails.
Definitions are one-line summaries written for the video, so they trade detail for speed. The video walks through each one with examples. More tutorials are on the videos page.