AI Platform Masterclass · Tier 0 · start here
Ten words, and you’re ready.
Never built with AI before? Good, this is the right place to start. The lessons assume these ten words, and nothing more. Read them once, and every lesson will make sense. No maths, no jargon, one plain sentence each.
Model (LLM)
A program that predicts the next word, trained on a huge amount of text. That is genuinely it. Everything else is built on top of that one trick.
Give it "the sky is", it predicts "blue".
Token
The model reads and writes in small chunks called tokens, roughly three quarters of a word. You pay per token, both the ones you send and the ones it writes back.
"unbelievable" is about 3 tokens.
Context window
The model’s short-term memory for a single request, measured in tokens. Everything it needs to answer, your rules, the question, any documents, must fit inside it.
Like a desk: only so much fits before things fall off.
Prompt
The text you send the model. The system prompt sets the rules ("you are a bank assistant, never reveal SSNs"); the user prompt is the actual request.
System = the job description. User = the task.
Embedding
Turning text into a list of numbers so a computer can measure how similar in meaning two pieces of text are. This is what makes search by meaning possible.
"car" and "automobile" land close together.
RAG
Retrieval-augmented generation. Fetch the relevant documents first, put them in the context window, then ask. It is how a model answers questions about your data, not just its training.
Look it up, then answer, instead of guessing.
Agent
A model put in a loop that can take actions, look at the result, and decide what to do next, instead of answering in one shot.
Think, act, observe, repeat, until done.
Tool call
The model asking your code to do something real (look up an order, issue a refund, send an email) and getting the result back to reason over.
The model says "call check_balance"; your code runs it.
Guardrail
Deterministic checks around the model, not inside it, that decide what is allowed in and what is allowed out. The model cannot argue with them.
A bouncer for inputs and outputs.
Trace
A record of everything that happened inside one request: every step, tool call, token, and cost. It is how you see and debug what the model actually did.
A flight recorder for one answer.
That’s the whole vocabulary.
Every lesson starts with a toy you operate, not a wall of text, so you learn by doing. Start with prompt engineering, the most concrete one, and work along the row.