Answer key: Exercise 1, Ask well

Time: 12 minutes (3 for round 1, 4 for round 2, 5 to compare and discuss). Handout: Exercise 1.

What usually happens

On the November 3, 2026 check run (Claude Sonnet, each request three times):

• Round 1: Claude didn't make up numbers. All three times it wrote a paragraph full of blanks, like "[X] households" and "[measurable outcome, e.g., a 20% increase in households served]", and asked for the real details. The risk moves to the person: someone in a hurry fills the blanks from memory, or leaves the example "20% increase" in.
• Round 2: all three answers used only the report's numbers. One added a reason the report doesn't give: the new Saturday hours were "for families who cannot visit during the week." That's the kind of sentence pairs should underline. It may be true, but nobody checked.

Answers on the day can differ: claude.ai has its own settings, and models change. Watch for rounding ("over 4,000 people" for 3,912 is wrong; "nearly 4,000" is fine), arithmetic ("about 41 hours per volunteer" is right, but now someone has to check it), and effects the report doesn't claim ("ensuring no child went hungry").

Every number in the report

1,284 households · 3,912 people · 1,406 children · 61,450 pounds · 47 volunteers · 1,930 hours · August 2 · 212 Saturday households. Also July to September 2026. Anything else is new and needs a source.

Discussion prompts

• "Which answer would you send?" Most pick round 2, with edits. Ask what they'd still change.
• "If Claude had written '1,248 households', would you have caught it?" A swapped digit is the error a quick read misses. Checking number by number catches it.
• "Is it OK that Claude worked out 41 hours per volunteer?" It's correct, and it's useful, but you now have to check the math. Some groups decide to ask for "only numbers from the report".

If something goes wrong

• No internet or Claude is down: hand out the printed sample answers from the latest check run and do the "compare" step on paper.
• Round 1 comes back with real-looking numbers instead of blanks: even better for the lesson. Have pairs find where each number came from. (It can't have come from the pantry; Claude never saw the report.)