Here's how Week 0 goes
First, one required step: read and sign the Governance Acknowledgment below. It turns on your Claude seat, and it's where you agree to the data rules you'll follow to keep scholar information safe. Then the week is three things, in order: get into Claude, get one real win on your own work, and see what Claude actually is. That's the whole week. The signature, the sign-in, and the win need your hands; the rest you can skim. Nobody is watching, and nothing here is graded.
By the end of this course, you will be using AI effectively. You'll get more out of it than someone who never learned how it works, because you will understand how large language models (LLMs) work, how to use Claude and its features, and how AI can speed up the work already on your plate, supplement your thinking, and open up new ideas.
Sign the Governance Acknowledgment
One required step turns on your Claude seat: read DC CAP's data rules (the four tiers for what goes into Claude and what stays out) and check the box to say you've read them. It takes a few minutes, and it comes before your first real prompt. Everything else this week is yours to explore at your pace. This step is the one thing you have to do first.
Read and sign it →Open Claude
Claude lives in the Claude Desktop app on your computer. Not set up yet? The 2-minute get-in walkthrough shows you exactly how to sign in. Already in? Skip down to the win below.
Get into Claude →Get a win
To build your confidence with Claude, start with a simple, relevant prompt. Experiment with different ways of prompting toward the same goal. You will make mistakes. It doesn't break anything! What matters is that you try.
Read Claude's reply again. What did it decide for you that you never actually told it?
Show what we noticed →
You never told Claude which day each thing was on, so it filled the schedule in on its own. Handy as a starting point, but it guessed those days, and your real week might look nothing like them. Noticing invented details like that is exactly the skill this course builds, and the next section shows you why it happens.
Here's my week: [list what's on your plate]. Turn this into a tight Monday-morning brief I can skim — what should I focus on each day?
- Click Copy above.
- Open Claude and click the message box at the bottom.
- Paste it in (right-click → Paste, or Ctrl-V / ⌘-V).
- Type over the part in
[brackets]with your real week, then send.
The same ask, two ways: “Write an email about the deadline” vs “Warm reminder to a busy senior that the renewal is due next week, under 120 words, one clear next step.”
Why the second comes back better →
Same Claude. The second one tells it who the message is for, what matters, and what done looks like, so the draft comes back usable. Week 1 turns this into a simple five-part recipe you can reuse.
Scale it past an email and the gap gets wider. Here's the same idea on a real project, planning a scholar event across four campuses, briefed three ways. You'll learn the five-part recipe in Week 1; this is what it's worth.
Meet the machine
You just watched Claude do something useful, and if you pushed it, you saw it get something wrong, too. Both come from the same place. Here's what's actually happening when you talk to it.
Claude has no record of what she did. It chooses the most plausible next word from patterns it learned, then does it again, and again, one word at a time. (The numbers are made up to show the idea; Claude's real percentages would differ.)
A model is something built from examples to make predictions — a weather model reads today's conditions and predicts tomorrow's. A language model does that with words: give it some text, and it predicts what plausibly comes next. Large is the scale. Claude was built by reading an enormous amount of text and finding the patterns in how language fits together. Claude keeps no copy of that text. The patterns are what it carries forward.
So when you ask a question, Claude builds an answer one word at a time, as the most plausible continuation of your words. On its own it looks nothing up — it works from the patterns it learned and whatever you put in the conversation. That one mechanism is the whole personality:
- It writes fluently on almost anything, because it learned the patterns of almost everything.
- It can be confidently wrong (people call this a hallucination) because "most plausible" and "true" aren't the same thing. A made-up citation matches the pattern of a real one perfectly.
- It starts each conversation blank — the model holds no memory of you between chats unless memory is deliberately switched on.
That's enough to use Claude well. Fast, fluent, useful, and worth a check when the facts matter.
Say it back: if someone asks what a large language model is, you can borrow this — "It predicts the next word, over and over, from patterns it learned in a huge amount of writing."
Know the room
Today you only need one part — the box where you type. Everything below is here just so nothing on screen surprises you. Tap any card to read what it's for. The gold "today" cards are the parts you've already met; the rest you'll simply recognize until later weeks.
What you can do now
Actually using Claude and critically engaging with the output is a key first step! Looking ahead, we will explore what's going on under the hood and how to get exactly what you need when working with Claude.
This is your own check that the win landed. It stays in your browser — nobody else sees it.
Under the hood
Skip it and lose nothing. Here for the curious — a plain look at how the prediction actually happens.
How does Claude predict, exactly?
Claude reads your text as a sequence and, for each next word, weighs how the patterns it learned fit what's there so far. It picks a likely word, adds it, and repeats — which is why answers arrive smoothly, left to right. (Technically it works in tokens — usually a piece of a word rather than a whole one — but "next word" is the right picture to keep.)
It holds the whole conversation in a working space called the context window. Early on there's room; in a very long chat the window fills, and the model can lose track of something you said far back. That's why a fresh, focused conversation often works better than one giant thread — something you'll use on purpose in later weeks.
If it's just predicting, why is it ever right?
Because true things appear together in writing far more often than false ones, so "most plausible" lands on "correct" a great deal of the time — especially for language work like drafting, summarizing, and reshaping. The gap shows up on specific facts it was never reliably taught: names, numbers, citations. That's the line you'll learn to check.
Why doesn't it know recent events, or my world?
Claude learned from an enormous amount of text up to a cutoff date. Anything after that, or anything specific to your work, was never in what it read, so it has no pattern for it. Ask it "what did my team decide yesterday?" and it has to guess. Knowing where its knowledge stops is half of knowing when to check it. One exception you'll see: when web search is on and Claude decides to look something up, it can read current pages, and it shows you what it found. The habit stays the same: know when it's working from patterns and when it's working from sources.