Here's how Week 1 goes
Week 0 showed you a machine that predicts from patterns. This week covers the other half: the context you put in front of it, which is the half you control.
What Claude sees
Picture the folder you'd hand that eager intern from Week 0. Whatever is in the folder is the whole briefing; whatever stayed on your desk may as well not exist.
Toggle items in and out of the folder below and watch what comes back change. Then try the two at the bottom that most people assume are in there.
Your folder: tap to add or remove
- Claude's standing instructions (always here)
Tap items on the left. The window and the answer below update as you go.
If it's not in the window, Claude doesn't know it.
Two kinds of knowing
Claude draws on exactly two sources. Telling them apart is what separates people who use AI well from people who get burned by it.
Patterns from training
General knowledge and fluent writing, fixed at a cutoff date. No current or reliable DC CAP specifics: it may have absorbed stale public facts about us, but it doesn't know our deadlines, our numbers, or how we write, and anything it seems to know needs checking.
The context you load
Your situation, your files, your instructions. Current, specific, and entirely under your control.
A hallucination is usually a prediction running without context. The fix is usually the same: put the missing thing in the window. (Claude can also misread what you gave it — which is why you check the draft against the file either way.)
You ask: "Summarize our partner retention results for the board." You attach nothing. What comes back reads beautifully. What's wrong with it?
Show what we noticed →
Every number in it is invented. Claude had patterns for "board summary about retention" and zero context about DC CAP's actual results, so it predicted plausible numbers. The same ask with the real one-pager attached comes back grounded in the file. Same Claude, different window.
The gate: what may enter
Governance is the check you run before anything goes in the folder, and you already signed the rules: the Governance Acknowledgment from Week 0. In the weekly flow it comes down to one question: could this point to one scholar or family?
The habit: describe the situation; the names and numbers stay in your files. Claude needs the situation, never the record.
The same task, two ways: re-engage a scholar who's gone quiet. Watch what each version puts into the window. (The scholar details below are invented for this example.)
"Maria G., GPA 2.1, SSN on file, hasn't logged in since Oct 3. Write her a re-engagement email."
"I have a first-gen sophomore who's missed our last two check-ins and seems at risk of not re-enrolling. Draft a warm, low-pressure email inviting them to reconnect."
Both get you a strong draft. Only one put a scholar's record into the window.
Show why it matters →
The safe version gave Claude the situation, which is all the prediction needed. Your read of the scholar's situation is what made the email work; the name and the GPA only added risk. The judgment stays with you, the situation goes in the window, the record stays in your files.
You took the names off a small campus breakdown. Safe to share now?
Show the answer →
Not if the group is small. Outcomes for six scholars at one campus point back to real people even with no names attached, because anyone who knows that campus can work out who's who. When a group is small enough to re-identify, describe the pattern instead of sharing the rows. The data rules you signed cover this and every other case; two minutes to reread whenever you're unsure.
And one thing to know before you ever need it. If something identifying a scholar ever lands in a chat by accident:
- Delete that conversation right away.
- Tell the AI Governance Lead or Co-Owner — Preston Magouirk or Angela Cammack — within two hours. They take it from there.
Pack the window
A good prompt is a packing list: it puts into the window exactly what Claude needs, and nothing it doesn't. Five parts, straight from Week 0's chart.
Files are the other half of packing: a grounded file beats a beautiful guess every time. Two habits when you attach:
- Open the file first and check what's in it. Nothing that points to one scholar or family goes in; describe those situations instead.
- Tell Claude what to do with it. "Pull the three numbers a funder would ask about" gets more than attaching in silence.
Team: the numbers are yours. 85% of the class of 2024 came back for year two, and seven partner cohorts have reached 100% retention in recent years. That's coaching showing up in the data. Renewal season is the next push, so let's get ahead of it this month. I'll bring the details Friday.
[Your name]
Every part of the window did work, and nothing in it points at any one scholar. (Those two results are real DC CAP figures, unlike the invented examples elsewhere on this page.)
[Attach one safe file; check what's in it first.] [Role: who Claude should be.] [Context: your situation in general terms, no names or records.] [Task: exactly what you want from the file, plus any limits.] [Example: one line showing the tone.] [Format: what done looks like.]
- Pick one real task that touches a document: a summary you owe, a draft to tighten, numbers to explain.
- Open the document and check it: nothing pointing to one scholar or family.
- Click Copy above, paste into Claude, and fill the brackets. Attach the file with the + at the left of the message box (or drag it in), then send.
- Read the draft against what you gave it. Something off? Add the missing context and send again.
Context that sticks
When a chat ends, its window empties, and memory may carry a few highlights forward without you picking which. Settings has a field called Instructions for Claude; whatever you write there rides along in every new chat, guaranteed.
I'm a Success Coach at DC CAP, a college-access nonprofit in Washington, DC. I work with first-generation college students at our partner universities. Keep drafts warm and plain, under 200 words unless I ask for more, and give me one clear next step. Never use real scholar names, even if I slip and include one.
Your instructions enter every window, so the gate applies to them too.
Show the two rules →
First, describe your role and audience, never specific scholars: standing context is no place for anything that identifies a scholar. Second, you can make Claude part of your own safety net: a line like "never use real scholar names" means the window itself pushes back if a name slips in.
Set yours now; three lines is plenty
In Claude, click your initials in the corner, choose Settings, and open Instructions for Claude. Write your role, your usual audience, and one taste. Every chat for the rest of this course starts smarter.
Your instructions are the smallest member of a family you'll build all course long: write the context once, and stop paying for it every morning:
What you can do now
You can see the window, guard it, and pack it. The four habits you just practiced have names: Anthropic's 4Ds, the framework this program runs on:
Delegation
You chose what to hand over: the task, and the folder that travels with it.
Description
You wrote the window: situation, task, example, format.
Discernment
You read the answer against what Claude actually had, and spotted where it guessed.
Diligence
You ran the gate: checked everything before it entered, and nothing identifying a scholar got through.
This is your own check that the week landed. It stays in your browser — nobody else sees it.
Every week stays open — move at your pace.
Keep the daily reps going: one real task a day, packed the way you practiced. Stuck on one? Bring it to Preston's office hours, 9–10 AM ET any weekday.
Under the hood
Skip it and lose nothing. Four questions people usually ask about the window.
Doesn't Claude have memory?
Where memory is available in our workspace, Claude keeps short notes about your recurring work and may bring some into a new chat. Treat it like a colleague's recollection of last week: welcome when it surfaces, too spotty to build on. The window is the only context you fully control, so anything that matters goes in by your hand: attach it, paste it, or write it into your instructions.
How big is the window?
Large (hundreds of pages of text) but finite. A long conversation slowly fills it, and details from far back can fade from Claude's attention even before the window is technically full. That's why a fresh, focused chat often beats one giant thread, and Week 2 turns that into a set of working habits: when to continue, when to start over, and how to carry the important parts forward.
What exactly is the system prompt?
The standing instructions Anthropic places at the front of every window: how Claude should behave, what tools it can use, where its limits are. You don't see it in the chat (Anthropic publishes claude.ai's), but it's why Claude has consistent manners, and your instructions ride in the window alongside it.
Where do my files and chats go?
DC CAP's Claude Enterprise workspace is set up so your inputs aren't used to train the model, and activity is auditable under the governance standards you signed. A personal AI account has none of that, and consumer tools may use what you type to train, so DC CAP work stays in the Enterprise workspace.