Not how it was built — there's no maths here and no diagrams. Just enough of a working model that the next seven lessons have something to hang off.
You already use Claude. You've had answers that were excellent and answers that were confidently, specifically wrong, and probably no way to tell which was coming. That's not a gap in your technique. It's a gap in your model — and you can't debug a system you have no model of. So we start here, and everything after this is a consequence.
Three consequences, in fact. They do almost all the work for the rest of the course, and each one ends with something you can actually do differently tomorrow.
Claude produces an answer sequentially — a chunk of text at a time, each one shaped by everything already in front of it: your question, the conversation so far, anything you attached, and the words it has already written.
That's the whole starting picture. Hold onto the word sequentially, because two of the three consequences fall directly out of it.
Ask a question with no document attached and no web search in play, and Claude produces an answer from what it learned. It is not opening a record and reading you the contents.
Here is what that can look like. Read these five references and decide, before scrolling, which ones you'd be willing to cite in something with your name on it.
"What are the key papers on the effect of open-plan offices on measured productivity?"
They are formatted perfectly. Volumes, issues, page ranges, plausible author names, real journals. Nothing on the page separates them. Here is what they actually are:
| 1 & 2 | Real. Bernstein & Turban is genuinely in Phil. Trans. R. Soc. B 373(1753); Kim & de Dear is genuinely in J. Environ. Psychol. 36, 18–26. Both check out exactly as printed. |
| 3 | The dangerous one. Those three researchers are real, they genuinely work on office noise and cognitive performance, and they genuinely published together in 2020 — but not under that title, not in that journal, and not at that volume and page. Real people, real subject, wrong paper. |
| 4 & 5 | Invented. No Whitmore & Castellano 2019. No Lindqvist & Aaron 2021. The journals are real; the papers are not. |
Item 3 is the one worth your attention, because it is the shape this failure usually takes in practice. Not a wild invention you'd catch at a glance — a near-miss, assembled from things that are individually true. The authors check out. The topic checks out. Only the specific artefact doesn't exist. If you verified that citation by searching the author names, you would find them, nod, and move on.
Almost everyone carries a quiet assumption that it's looking things up — that somewhere behind the reply is a record being consulted, and mistakes are retrieval errors. Drop that. Unaided, there is no lookup, and the confident specificity of a fabricated citation is not a malfunction. It's the same machinery that writes your emails well, running with nothing to check itself against.
What to do about it: when it matters, give it something to work from — attach the document, turn on search — or ask it to quote the line it used. That single habit is most of Lesson 5.
Because the answer comes out in order, it's tempting to conclude that the words are the thinking, and that whatever isn't written down didn't happen. That's not right either, and it's worth being precise, because the practical advice survives while the story doesn't.
Work can happen that the final answer never shows you. And when Claude does narrate its reasoning, that narration isn't guaranteed to be a complete account of what produced the answer. So don't treat the visible words as a log of the machinery.
What is reliably true, and useful:
You don't need a theory of why to use it. A long multi-step calculation asked bare, and the same calculation asked with "work through it one step at a time, showing each step," are not the same request, and the second one tends to hold up better. Lesson 7 covers the setting that does this for you.
Claude learned from text up to a point, and knows nothing after it. That part most people already half-know. Here's the part that catches them:
It generally can't tell that it doesn't know. A question about something recent doesn't feel different to it than a question about something a decade old, so the answer arrives in exactly the same confident register. There's no flicker, no hedge, no tell.
Asking Claude when its own information ends is unreliable — it's answering that question the same way it answers any other. So the answer you get is not a fact about the system; it's another generated claim. If currency matters, the repair isn't to ask. It's to give it a source, or turn search on.
Which is worth sitting with for a second, because it means the two failure modes in this lesson — never knew it and stopped learning — look identical from the outside. Same confident tone, same specificity, same absence of any warning. One repair covers both.
Ask the same question twice in two fresh chats and you can get two different answers. Not contradictory usually, but not identical either.
That matters more than it sounds, and here's why: it means regenerating an unchanged prompt is not a repair. You're drawing again from the same place. If the prompt was the problem, the second draw has the same problem. Lesson 4 gives that its own name — the reroll trap — because it's the single most common wasted action by people who use Claude casually.
Four small runs. None of them takes long, and none of them has a required outcome — write down what you predict first, then record what actually happened, including "it was fine." A run that comes out better than the lesson implies is information, not a broken exercise. These systems improve, which is the entire reason the skill being trained here is checking rather than assuming.
1 · The sources run. Pick a topic narrow enough to be obscure and that you could verify. Ask for five references, web search off. Then check two of them. Then ask again with search on and compare what changes.
2 · The recency run. Ask about something from the last few weeks, search off, then on. Notice whether the first answer signalled any uncertainty.
3 · The room-to-work run. Take a genuinely fiddly multi-step calculation. Ask it bare. Then ask it again in a fresh chat with "work through it one step at a time." Compare.
4 · The twice run. Same question, two fresh chats. Note how much drifts.
Four boxes. Tap as you go — it saves in this browser only.
From memory, without scrolling up: name the three consequences from this lesson, and the one repair that covers two of them.
One: with no source, it's generating, not looking up. Two: sequential output isn't a log of the reasoning — but giving it room to work helps. Three: what it learned has an end date, and it can't tell that it doesn't know. The shared repair for one and three: give it a source, or turn search on. Asking the system about itself is not the check.
No source → generating, not looking up.
Confident and specific is not evidence of true.
It can't tell you what it doesn't know.
Anthropic, AI Fluency: Framework and Foundations — a free twelve-module course built with Prof. Joseph Feller (University College Cork) and Prof. Rick Dakan (Ringling College). anthropic.com/learn/claude-for-you
Watch the Delegation and Description sections. It's the closest thing to a shared vocabulary in this space — its four Ds (Delegation, Description, Discernment, Diligence) are terms you'll meet elsewhere, and its Description–Discernment loop is more or less what Lessons 3 and 4 build out in detail. Worth an hour before you go further.