Facilitator guide

You don't need to be technical to run this. You need to have used Claude a few times, and to be comfortable saying "I don't know, let's check." That's the whole workshop in one sentence.

The hour

A week before

1. Ask about the organization's AI policy. If there is one, read it and bring it. If it conflicts with anything here, the policy wins; say so in the room.
2. Check accounts. Each pair needs one laptop signed in to Claude (a free account works). If people can't sign in, pairs can share yours, or use the printed sample answers.
3. Run the check (or ask the kit's maintainer to): uv run workshop check. It takes a few minutes and costs about ten cents. Read the summary, and print a few answers from checks/results/<date>/answers.jsonl as paper backups for Exercises 1 and 2.
4. Check the SNAP number in the Exercise 2 key is still current (it changes every October 1; see the RUNBOOK).
5. Print, for 10 people (5 pairs):

Block by block

0–5: Welcome and pre-quiz

Say: "This hour is about four habits for using Claude at work: asking well, checking, keeping clients safe, and knowing when not to use it. It's not about becoming an expert."

Hand out the quiz. "This isn't graded and has no names. We'll take it again at the end, so we can tell if the hour was worth it." Give 3 minutes. Collect them face down.

5–12: What Claude is and isn't

Use slides 4 to 6. The points to land:

• Claude writes by predicting likely words. That's why it's fluent, and why it can be fluent and wrong at the same time.
• It learned from text up to a date. It doesn't know this month's rules, your organization, or your clients, unless you tell it, and you shouldn't tell it about clients.
• It's good at drafts, rewording, summaries and ideas. It's not a source of facts.

Ask: "Who has used it already? What for?" Take two or three answers; don't let this run long.

12–24: Exercise 1, Ask well

Pairs, one handout each. Keep time: round 1 at minute 15, round 2 at minute 19, compare at 24. Use the key for what to expect and the discussion prompts.

24–36: Exercise 2, Check before using

The SNAP number is real and current; the Tennessee conflict is real too. Spend the last 5 minutes on step 3. The answer you want from the room: "We don't know, so we tell her to apply and we call the agency or 211." Key.

36–46: Exercise 3, Protect clients

Start by saying nobody pastes the emails into Claude. When pairs try their de-identified request, point out the reply is nearly as useful. Hand out the privacy sheet at the end. Key.

On the check run, a reply to the pasted email used the client's name and her child's name and mentioned the breathing machine; the de-identified request got a template with none of that. That's the trade: a more personal draft, for a client's details sitting in a chat.

46–55: Exercise 4, When not to use it

Groups of three or four. Five minutes to sort, four to discuss. Ask each group for the card they argued about most. Key.

55–60: Close

Post-quiz (3 minutes), then hand out the takeaway and feedback form. Thank them. Tell them who to ask about the organization's AI policy.

If you're running behind

• Cut Exercise 4's discussion to the two hardest cards (3 and 8).
• Don't cut Exercise 3 or the post-quiz.

After

1. Tally the quizzes with the tally sheet: count correct answers per question, before and after.
2. Read the feedback forms.
3. Send the tallies and a summary of the feedback (no names) to the kit's maintainer, or add them to the repo under pilots/. That's how this kit gets evidence that it works.

Running it in Spanish

Everything attendees see is also in Spanish: the handouts, the privacy sheet, the quiz, the takeaway, the feedback form and the slides. The privacy sheet quotes Anthropic's own Spanish Privacy Center pages. This guide and the answer keys stay in English.

What Claude did with the Spanish prompts on the November 4, 2026 check (three runs each) matched English: blanks in round 1 of Exercise 1 and only the report's numbers in round 2; for SNAP, last year's $3,483 with $3,575 called an estimate (2 of 3) and a pointer to the USDA or the state agency (3 of 3); a reply to the pasted email used the client's and her child's names. Things to know:

• The USDA page is in English. Help pairs find the "Gross Monthly Income Limit" table and the row for 4.
• The news quote in Exercise 2 is translated, and the handout says so.
• The Spanish was written with Claude's help and hasn't been reviewed by a native speaker. Have someone fluent read the handouts before the first Spanish session, and fix anything that reads awkwardly.

What people ask

• "Is Claude listening to everything?" Point to the privacy sheet: it quotes what Anthropic says, with dates and links. Don't add to it from memory.
• "Is it going to take my job?" Be honest that you don't know. Today's point is narrower: if you use it, use it carefully.
• "Which is better, Claude or another AI tool?" Out of scope today. The four habits apply to all of them.