← BACK TO THE HUB
✓ LEVEL CLEARED
THE FUNDAMENTALS · PILLAR 05

The AI glossary

26 terms you keep running into, each decoded in one plain sentence. No circular definitions, no "you should already know this." Bookmark it and check back whenever a word trips you up, then go deeper on the big ones in What is an AI agent and Skills.

THE BASICS
Model (LLM)
The engine at the center of everything, a large language model trained to predict the most useful next words; Claude and GPT are models.
Agent
A program that uses a model to decide what to do next and then actually does it, in a loop, until a job is finished.
Chatbot
A model wrapped in a chat window that remembers the current conversation but cannot take actions outside it.
Prompt
The text you give a model; the clearer and more specific it is, the better the answer.
System prompt (soul file)
The standing instructions that set who an agent is, its name, voice, job, and rules, before any user even talks to it.
Skill
A single file that teaches an agent one job done well; stack many and the agent grows more capable.
Token
The unit models read and write in, roughly a word-piece; pricing and limits are counted in tokens, not words.
Context window
The amount of text a model can hold in mind at once, everything past it is forgotten unless stored elsewhere.
HOW THEY BEHAVE
Hallucination
When a model states something false with total confidence, because it is predicting plausible text, not looking up facts.
Temperature
A dial from predictable to creative; low temperature gives safe, repeatable answers, high gives varied, riskier ones.
Fine-tuning
Further training a base model on your own examples so it specializes, as opposed to just prompting it well.
Parameters (weights)
The billions of internal numbers a model learned during training; more is not always better.
Grounding
Giving a model real source material to answer from so it stops guessing and sticks to the facts you provided.
Prompt injection
An attack where hidden text tricks an agent into ignoring its instructions; the reason agents should not blindly trust web content.
MAKING THEM USEFUL
RAG
Retrieval-augmented generation: the agent looks something up in a knowledge source first, then answers from what it found.
Tool use (function calling)
How a model triggers real actions, it names a tool and the inputs, and the surrounding program runs it and reports back.
MCP
Model Context Protocol, a shared standard for plugging tools and data sources into an agent so they all connect the same way.
Memory
How context is handed back to the model each turn, short-term is the current chat, long-term is saved outside the model and pulled in when relevant.
Embedding
A way of turning text into numbers so a computer can find things by meaning instead of exact wording.
Vector database
Where embeddings are stored so an agent can search a big pile of documents by meaning, the engine behind most RAG.
Agentic loop
The think, act, observe, repeat cycle that lets an agent take one step at a time until the goal is met.
THE ECOSYSTEM
API
The doorway that lets your software talk to a model or tool programmatically, instead of through a chat box.
Multimodal
A model that handles more than text, images, audio, or video, in and out.
Open-weight vs closed
Whether a model can be downloaded and run yourself (open-weight) or only used through the provider (closed).
Guardrails
Rules and checks that keep an agent from doing something unsafe, off-brand, or off-topic.
Latency
How long the model takes to respond, the lag between asking and getting an answer.
Know the words? Build the thing.

The glossary is the vocabulary. Now put it together into a real agent.

BUILD AN AGENT →