Learning objectives. By the end of this module you will be able to explain, in plain language, what an AI model actually is and isn't; distinguish a question, a prompt, an instruction, a workflow, and an agent; name the two most common failure modes (hallucination and context loss) and why they happen; and know exactly where this course's scope ends and the Holistic Agent course's begins.

Why this matters

Most disappointing AI experiences trace back to a mismatch between what someone assumed they were talking to and what they actually were. If you think you're talking to a system that remembers everything, checks its facts, and understands your business the way a colleague does, ordinary AI behaviour will look like a string of failures. If you understand what you're actually working with, the same behaviour looks like exactly what to expect — and you can work around it deliberately instead of being surprised by it repeatedly.

Plain-English explanation

An AI model — specifically, a large language model (LLM), the kind behind ChatGPT, Claude, and Gemini — is a system trained on enormous amounts of text to predict what word (technically, what token — a fragment of a word) comes next, given everything before it. That single sentence explains a surprising amount of both its power and its limitations.

It is extremely good at producing fluent, well-structured, contextually appropriate text, because it has seen an enormous number of examples of what fluent, well-structured, contextually appropriate text looks like. It is not, by itself, a database, a calculator, a fact-checker, or a system with memory of you between separate conversations. Everything that feels like memory in a product you use — Claude, ChatGPT, or otherwise — is engineering built around the model: your previous messages are being re-sent to it every single time, or a separate memory system is retrieving facts about you and inserting them before your question is even seen.

Core lesson

PLAIN ENGLISH — five things people confuse

Term	What it actually is	Example
A question	Something you'd ask a knowledgeable friend. One-shot, no real structure needed.	"What's the capital of Peru?"
A prompt	A more deliberate request — you're specifying the task, not just asking a question.	"Write a two-paragraph LinkedIn post announcing our new product, confident but not salesy, aimed at small business owners."
An instruction	A prompt with authority behind it, usually persistent — shapes every response an AI assistant gives, not just one.	A system prompt or custom instructions.
A workflow	A defined sequence where AI may do one or more steps. Every step is known in advance — this is AI automation.	Receive an email, draft a reply, wait for approval, send it.
An agent	Given an objective, works out its own steps using tools to achieve it.	Reads the email, checks your calendar, drafts a reply referencing a time you're free, asks you to approve before sending.

This course lives almost entirely in the first three. Module 10 opens the door into the fourth and fifth — and the companion Holistic Agent course is where you go to walk all the way through it.

PRACTITIONER — what actually goes wrong, and why

Hallucination. The model produces fluent, confident, incorrect output. This is not a bug that will be patched out — it's structural. The model is optimised to produce a plausible continuation of your text, and a plausible-sounding case reference, statistic, or citation is exactly what "a plausible continuation" looks like when the model doesn't actually have the real one. You cannot out-word this problem with clever phrasing. You can design around it — covered fully in Modules 5, 6, and 11.

Context loss / no persistent memory. Each time you send a message, the model has no innate memory of the last one — the product you're using resends the conversation, or a separate system injects remembered facts. This is why a very long conversation can start behaving oddly: not because the AI is "getting tired," but because the growing context is genuinely harder for it to use well (Module 5 covers exactly why).

Confident uncertainty. A model's stated confidence — "I'm quite sure that..." — carries very little information about whether it's actually correct. This is one of the most consequential things to internalise early: treat fluency and confidence as style, not as evidence.

ADVANCED — what's actually happening under the hood

Models process tokens, not words — "unbelievable" might be three tokens. Cost, speed, and context limits are all really the same conversation, denominated in tokens. Each response is one inference — a single, stateless pass. The model's general knowledge was fixed at training time (its training cutoff); it doesn't know today's date, your prices, or last week's news unless something puts that information in front of it. Modern models are increasingly multimodal — they can process images, PDFs, and audio, not just text — and some can spend extra computation "thinking" before answering (a reasoning model), which genuinely improves multi-step accuracy at a real cost in tokens and latency, not a free upgrade (Module 4 has the research on when this actually helps).

Beginner example

Weak: "Tell me about marketing." Better: "Explain the difference between paid and organic marketing to someone running a five-person bakery, in plain language, with one example of each."

The second version isn't longer because more is always better — it's better because it answers a specific person's specific situation instead of a generic textbook entry.

Business example

A small accountancy practice used AI as a general Q&A tool for the first year — "what's a good way to explain VAT to a client?" — and found it useful but unremarkable. Once the office manager understood the distinction between a question and an instruction, she built a standing prompt: "You are drafting client-facing emails for a UK accountancy practice. Plain English, warm but professional, always end with a clear next step. Flag anything that sounds like it's giving financial advice rather than just factual explanation." The same underlying tool became measurably more useful — not because the AI changed, but because the instruction layer around it did.

Practical exercise

Take the last three things you asked an AI to do. For each, classify it: was it really a question, a prompt, an instruction, or something you were hoping would behave like an agent? For any you classify as "hoping it would behave like an agent," rewrite it as an explicit prompt instead, and note what changed.

Common mistakes

Assuming the AI remembers your last conversation because the product felt like it did.
Treating confident phrasing as evidence of accuracy.
Expecting an assistant to check facts, calculate exactly, or know today's date unless you've told it to use a tool that can.
Calling every AI feature you encounter "an agent" — which makes it impossible to reason sensibly about what it can and can't be trusted to do (the Holistic Agent course opens with exactly this same warning, for exactly this reason).

Expert insight

The single biggest jump in usefulness most people experience with AI has nothing to do with a cleverer prompt. It comes from correctly recalibrating what they're talking to — from "an oracle that knows things" to "a fluent reasoning system that needs to be told what it needs to know." Everything from Module 2 onward is really just that recalibration, made systematic.

Knowledge check

Why does an AI model have no memory between separate conversations, even in a product that seems to remember you?
What is hallucination, and why is it structural rather than a bug to be fixed?
Distinguish a prompt, an instruction, and an agent, with one example of each from your own work.
Why does a model's stated confidence carry little information about its accuracy?
Name one thing an AI model cannot do without a tool, even though it might attempt to answer anyway.

Module summary

An AI model predicts plausible continuations of text; it does not remember, verify, or calculate by default. The question/prompt/instruction/workflow/agent ladder is the tool for keeping your expectations — and your requests — calibrated to what you're actually working with. This course lives mainly in the first three rungs; Module 10 opens the door to the rest.

Further exploration

Anthropic, Effective Context Engineering for AI Agents (2025) — introduced properly in Module 5, but worth knowing it exists from here.

Understanding How to Work With AI