# Pre-work — AI for Organizations (Executive Intensive)

**Time required:** ~1 hour. **Do this before Session 1 (Monday).**

You will use AI tools *with your own hands* throughout this week. The point of the
course is tangible, working knowledge — not a tour of screenshots. The setup below
makes sure you can build alongside the lectures.

---

## 1. Get access to Claude (15 min)

You need a working Claude account before Monday.

- **Claude (web/desktop)** — sign in at [claude.ai](https://claude.ai). This is enough
  for Assignments 1 and 2.
- **Claude Code (optional, recommended)** — the command-line "harness" we study in
  Session 3. If you're comfortable in a terminal, install it; if not, you can complete
  every assignment in the web app, and the AI tutor will help you adapt.

> If your organization restricts external AI tools, flag this to the course team now.
> We can provide a sandboxed environment so you are never asked to paste anything
> sensitive.

## 2. Confirm you can hold a long conversation (10 min)

Open Claude and have a real back-and-forth about something you know well — a project,
a market, a decision. Keep going for 15–20 exchanges. You don't need to do anything
with it yet; you're just getting a feel for the rhythm. In Session 1 we'll explain why
that conversation gets *worse* if you let it run long enough — and what to do about it.

## 3. Read the primer (20 min)

Read the short primer below. It introduces the four ideas Session 1 builds on. You are
not expected to master them — just to have seen the words once.

## 4. Bring a real problem (5 min)

Pick one repeatable task from your own work — something you or your team does the same
way every week (a weekly report, a briefing, a review checklist, an onboarding email).
You'll turn it into a *skill* in Assignment 2 and reference it all week. Write it down.

---

## Primer: four ideas we'll make concrete

**1. The context window is finite.**
A language model can only "see" a limited amount of text at once — its context window.
Everything in the conversation, plus any documents and instructions, shares that budget.
Performance starts to degrade *well before* the window is technically full. This single
constraint explains an enormous amount of AI behavior.

**2. Not all positions in the window are equal.**
Models reliably recall what's at the **beginning** and **end** of their context far
better than what's stuck in the **middle** — a U-shaped curve documented across models
("Lost in the Middle," Liu et al., 2023). Where you put information matters as much as
whether you include it.

**3. A raw model is a pure function: text in, text out.**
On its own, a model changes nothing in the world. It reads tokens and writes tokens.
The interesting — and risky — work happens at the **boundary** where something lets the
model touch files, run commands, or send messages. Those are *side effects*.

**4. Harnesses, skills, and agents.**
A **harness** is the software that gives a model hands — the ability to act. A **skill**
*adds* context: instructions and know-how loaded into the window so the model can do
something specific. An **agent** *subtracts* context: a fresh worker with a clean window,
handed one scoped task. Same model — opposite moves on the context window.

---

## Checklist before Monday

- [ ] I can sign in to Claude.
- [ ] (Optional) Claude Code is installed.
- [ ] I've had one long practice conversation.
- [ ] I've read the primer.
- [ ] I've written down one repeatable task from my work.

See you in Session 1.