How AI Models Work

1What an AI Model Actually Is

AI does not think. It predicts patterns at superhuman speed. That distinction changes how you use it — once you understand what's actually happening, you can give the AI the right inputs to get the right outputs. Pick whichever of these three analogies clicks, and keep it in mind through every lesson that follows.

Analogy	What It Means	What It Tells You
The well-read intern	Processed billions of documents, but predicts text rather than understanding it	Vague instructions produce vague results
The limited desk	The "context window" is its working memory — when it fills up, older information drops off	Paste only what the AI needs, not everything you have
The literal executor	Follows instructions exactly as written; does not infer what you meant	Be specific — it won't fill gaps the way a colleague would

2The Terms You Need to Know

You'll see these words constantly — in Claude, in this course, and in conversations about AI. You don't need to memorize them; you need to recognize them. Start with the four marked below — you'll use these every day.

Term	What It Means	Why You Care
Prompt *	The instruction or question you type to the AI	The quality of your prompt determines the quality of your answer
Context window *	The AI's working memory — how much text it can process in one conversation	Paste too much and it forgets earlier parts of the conversation
Token	A chunk of text, roughly three-quarters of a word	Large files cost more tokens and can overflow the context window
Model	The specific AI you're talking to	Different models have different speeds, costs, and capabilities
LLM	Large Language Model — the technical name for AI like Claude	The technology running behind Claude and similar tools
Generative AI	AI that creates new content rather than sorting existing data	Claude is a generative AI tool
MCP Server *	A live connection between your AI assistant and another system	Lets your agent query real data instead of guessing
Agent *	A pre-configured AI assistant with specific instructions and connections	The building block behind most real AI workflows
Knowledge Base	A collection of documents the AI can search to answer questions	Upload your team's SOPs so the AI can answer from them
RAG	Retrieval Augmented Generation — fetch relevant documents first, then answer	This is how Knowledge Bases work behind the scenes

3Three Stages of AI Capability

Not all AI tools behave the same way — there's a spectrum from simple text generation to fully autonomous multi-agent systems. Knowing where a tool sits tells you what you can ask it to do, and what you still need to do yourself.

Stage 1: AI Assistant	Stage 2: AI Agent	Stage 3: Agentic AI System
Follows rules. Automates simple tasks. You provide all context.	Achieves goals. Handles multi-step workflows. Pulls live data from connected systems.	Fully autonomous. Multiple specialized agents work together and consolidate results.
Example: paste a transcript into Claude, ask it to summarize. It does exactly that and stops.	Example: ask an agent to review 40 résumés against a job description and rank the top candidates. It reads each, compares, and returns a ranked list.	Example: a coordinating agent delegates to separate screening, scheduling, and drafting agents, then delivers one consolidated update.

4How Agents Actually Work

When you send a message to an AI agent, more happens than you see. It reasons, selects tools, queries live systems, and loops back to check its own work before answering. This loop is what makes agents powerful — and it's how you debug a wrong answer.

1. You type a prompt. "Which candidates have been in 'Interview Scheduled' for more than 10 days?"
2. The agent receives it and reads its own standing instructions.
3. The agent invokes the model, which reasons about what to do next.
4. The model selects a tool — the right source and the right query.
5. The tool executes against live data and returns results.
6. The agent reasons about the results, looping back if it needs more.
7. The agent returns the final response — one consolidated, readable answer.

Troubleshooting: what if the agent gives a wrong answer?

The agent returns outdated or incorrect data. Cause: a connected tool returned stale data, or the agent picked the wrong one. Fix: ask "Walk me through how you got that answer" — this surfaces which tool it used.

The agent says it can't answer. Cause: the required connection isn't enabled, or you lack permission on that system. Fix: check your AI tool's settings for the relevant connection first.

5What This Means for Your Work

Understanding how AI works changes how you use it. Three principles separate effective AI users from frustrated ones:

Verification

Related Resources

• claude.ai — The Claude AI platform
• Claude documentation — Official documentation
• Claude support — Help center and troubleshooting