11 · Teaching kit

A proposed module plan, demo scripts, exercises, the questions you will definitely be asked, and the misconceptions worth pre-empting. Take what's useful; the substance is in §01–§10.

Suggested module: 90 minutes

Pre-work

Assign roughly 80 minutes of Anthropic's own free courses before the session — AI capabilities and limitations (15 min), Claude 101 (1 hr), and the 5-minute Delegating your first task in Claude Cowork video. They arrive with shared vocabulary and having seen the interaction model, which buys back most of your first segment. Full rationale and the assignment table in §13.

If you only have 30 minutes

Keep: the frame (5), two products one engine (5), one live demo (10), the two-condition risk rule (5), the decision tree (5). Drop everything else. The demo is non-negotiable — abstract explanation of agentic work does not land.

If you have a half day

Add hands-on. Segment 3 below becomes 90 minutes of participants doing their own first task on their own data, with you circulating. That is where adoption actually happens.

Demo scripts

Demo 1 · Cowork — "the unglamorous win"

1. Show a messy folder of ~20 real-ish files: PDFs, spreadsheets, a few emails saved as text.
2. Create a project pointing at it. Add two lines of Instructions ("Always cite the source file for any number you quote. Use our fiscal quarters, not calendar.").
3. One prompt: "Work out what's in this folder, reconcile the numbers across the spreadsheets, and give me a two-page summary with a table of the discrepancies and where each one comes from."
4. Narrate while it works — point at the plan, the file reads, the permission prompts.
5. Open the output. Then open one of the source files and check a cited number live. Do this in front of them. Verification in public is the most persuasive thing you can do.

Demo 2 · Claude Code — "for the non-engineers in the room"

1. Open a terminal in a small repo. Show CLAUDE.md and read two lines of it aloud — "this is how the team writes down its conventions once."
2. Ask it to do something legible without code literacy: "What does this project do, and what would break first if traffic went up 100×?"
3. Then something concrete: "Add input validation to the signup endpoint, write a test for it, and run the tests."
4. Show Shift+Tab cycling permission modes. Show Esc Esc undoing an edit.
5. Land the point: this is the same agent that just built the summary; the difference is what it's pointed at.

Demo 3 · The bridge (only if the room is Pro/Max)

Brief Dispatch from your phone with something mixed — "look at last week's incident, write the customer-facing note, and open a PR adding the missing alert." Show the child tasks appearing in the sidebar and landing in different products. If your audience is on Enterprise seats, skip this and describe it instead.

Hands-on exercises

Questions you will be asked

Misconceptions to pre-empt

Address these before someone raises them; it reads as expertise rather than defensiveness.

1. "Cowork is Claude Code for normal people." Architecturally related, operationally very different. Different execution location, different context unit, no hooks, no git, no CI, no SDK, doesn't share local config.
2. "The desktop app means it's local." The opposite. Say this early and say it clearly.
3. "Claude Code is only for code." Its own docs disagree, and output styles exist to adapt it.
4. "More autonomy is better." Auto mode costs more and screens rather than blocks. Match the mode to the stakes.
5. "If it's in my instructions, it will happen." Instructions are context, not enforcement. This is a direct quote-able point from the docs.
6. "Enterprise plans get everything." Computer use and Dispatch are Pro/Max only.
7. "It runs in a sandbox so it's safe." Cowork's isolation protects your machine and network. It does not protect the connected accounts and folders you deliberately granted it.
8. "Prompt injection is theoretical." Anthropic names it as the central threat for Cowork and ships classifiers, action screening and RL training against it. Treat it as real and teach the two-condition rule.

Vocabulary sheet

Before each cohort — verification checklist

1. Re-read the Claude Desktop changelog for Cowork changes since your last session.
2. Re-read the latest Claude Code what's-new entry.
3. Re-check pricing on claude.com/pricing — including which seat type your audience holds.
4. Re-check feature availability — especially the beta and research-preview labels in §05.
5. Re-check the course catalogue at claude.com/resources/courses before sending pre-work — durations and titles change, and the hub page does not reflect the catalogue (§13).
6. Confirm the seat type of the people in the room, then confirm your demos will actually work on it.
7. Run both demos end to end the morning of. Agentic demos fail differently every time.