Using Claude at a Nonprofit
A ready-to-run, one-hour workshop that teaches nonprofit staff who don't write code to use Claude carefully: ask well, check answers before using them, keep client information out, and know when not to use it. Built so someone else can run it without me in the room.
The brief
Small nonprofits are adopting AI tools without training. A one-hour session, run by someone who isn't a technologist, can change the four habits that matter most. See it at nathanaelhub.github.io/nonprofit-ai-workshop.
The build
- Everything a facilitator needs: a minute-by-minute guide, slides with speaker notes, four exercises with answer keys, a one-page privacy sheet, a quiz to give before and after, a feedback form and a runbook for keeping it current.
- Real stakes in the exercises: the SNAP income limit and Tennessee's unresolved 200% question from my benefits screener; fictional intake emails full of private details to practice removing.
- Accuracy enforced by tests: every statement about how Anthropic handles chats is quoted from its Privacy Center and checked against a cached copy; attendee pages must read at grade 8 or below; the hour must add up to 60 minutes.
- A pre-workshop check: workshop check runs each exercise through Claude three times and scores the answers in code, so the facilitator knows what Claude says that week. About $0.10.
What I learned
My spec predicted Claude would invent numbers when asked vaguely for a grant paragraph. It didn't: every time, it left blanks like "[X] households" and asked for the real figures. The answer keys now say what actually happened, and the lesson moved to the person filling in the blanks. My first scoring of those answers was wrong too, until I read them. The kit hasn't been run with real staff yet, and says so.
Stack
Markdown and a small Python build, Claude (Sonnet through the Claude CLI), pytest, GitHub Pages.