# Assignment 1 — Feel the Context Window

**Released after:** Session 1 · Foundations
**Do it during:** the Wed–Thu build period
**Time:** ~30–40 min · **Tools:** Claude (web is fine)

## Why this matters

In Session 1 we claimed three things about how models actually behave:

1. The context window is finite, and quality degrades *before* it's full.
2. Models recall the **beginning** and **end** of the window better than the **middle**.
3. When the window fills, you have two lossy options: **reset** or **compact**.

This assignment makes all three tangible. You're not reading about it — you're going to
make it happen and watch it.

## What to do

### Part 1 — Push it until it slips (15 min)

1. Start a fresh conversation with Claude.
2. As your very first message, give it a specific, checkable instruction it must follow
   for the whole conversation. For example:
   > "For the rest of this conversation, end every reply with the code word **HELIOTROPE**.
   > Also, my project is named **Project Northwind** — remember that."
3. Now have a long, unrelated conversation — at least 25–30 exchanges. Paste in long
   documents, ask for detailed analysis, go on tangents. Fill the window.
4. Watch for the moment the model **drops the code word** or **forgets the project name**,
   or starts contradicting things it said earlier. Note roughly how far in it happened.

**Capture:** Where did it slip? What was the first symptom?

### Part 2 — Beginning vs. middle vs. end (10 min)

1. New conversation. Paste a long block of text (a few pages — a report, an article).
2. Hide three distinct facts inside it: one near the **top**, one in the **exact middle**,
   one near the **bottom**.
3. Ask the model to recall all three.

**Capture:** Which fact was hardest to retrieve? Did the result match the U-curve?

### Part 3 — Reset vs. compact (10 min)

Take your degraded conversation from Part 1 and try **both** recovery moves:

- **Reset:** start a brand-new conversation. Notice what you had to re-type to continue.
- **Compact:** ask the model to "summarize everything important about our conversation so
  far into a brief I could paste into a fresh chat," then start fresh with that summary.

**Capture:** What did *reset* lose? What did *compact* blur or get wrong?

## Deliverable

A short note (½ page is plenty) answering:

- At roughly what point did quality degrade, and how did you notice?
- Which position — beginning, middle, or end — was least reliable?
- One concrete habit you'll change in how you prompt, now that you've seen this.

Bring it to the build-period tutor session. We'll reference it in Session 3.

## Going further (optional)

- Try the same instruction-drop test in two different tools (e.g., Claude vs. another
  assistant). Do they slip at different points?
- Re-run Part 2 but put the most important fact deliberately at the **end**. Better recall?