Pin this at every desk

How to talk to it

The policy says what is allowed. This page is how the work actually gets good. Eight habits, learned in a week, used forever.

Work the loop

One prompt is never the job. The rhythm is: describe what you need, look hard at what came back, refine your description, go again. You are not accepting or rejecting drafts, you are building shared understanding, the way you would with a new colleague.

Generic in, generic out. “Tell me about Seattle housing policy” earns an answer exactly that vague.

Describe three things, not one

/The product: what you want, its format, its audience. “A one-page overview of Seattle housing assistance, organized by program, for a board that has ten minutes.”
/The process: how to get there. “Focus on the last two years. Compare to Portland, which we know. Cite official sources only.”
/The performance: how to behave. “Be a skeptical reviewer.” “Practical, not academic.” If you need a devil’s advocate or a hype man, ask for one.

Feed it your voice

Upload the past proposals that won, the mission one-pager, the report the board loved. Then tell it to learn your voice before writing a word. It cannot know your track record, your partnerships or your relationships, so you inject them.

Delegate the first draft. Never the final one. The last pass is always yours, and you should be able to stand behind every sentence.

Check before it leaves the building

Before anything goes to a donor, funder or client, three questions:

/Are the numbers, names and dates real? Verify every specific figure against a primary source.
/Is anything stale? Ask it: “confirm this from official government sites” and “show me the source for these deadlines.”
/Does it sound like us, and does it describe the people we serve the way they describe themselves?

For data: test it on last quarter

Before trusting AI with analysis, hand it a past dataset where you already know the answer. If it reproduces your known result, trust it on new data. If it cannot after a few rounds of refining, you have learned something better: that task stays with you.

Keep notes on what it missed. “Weight by program type” asked once is a lesson; asked every quarter it is your checklist.

Strip the names first

Analysis needs patterns, never people. Before a spreadsheet goes in, delete the name and contact columns. Keep zip codes if geography matters. Describe the situation, never the person, and the same data delivers the same insight with none of the risk.

If you overshare

/Delete the conversation from the history.
/For anything serious, request deletion from the provider (Claude: privacy.claude.com).
/Tell the policy owner the same day. Reporting is never punished.

Two questions that keep us honest

Can we explain what the AI is doing? If yes, that is healthy help. If no, rework the process until you can.

Would knowing AI was involved change how the reader sees this work? Then they need to know.

Homework for every seat: Anthropic and GivingTuesday publish a free course, “AI Fluency for nonprofits,” about one hour of video, built for exactly this. Send everyone through it in week one: anthropic.skilljar.com/ai-fluency-for-nonprofits

Applied AI Advisor