My AI donor advisor
A generic model emulates a generic funder. A model with your writing emulates you.
I started asking AI about political funding. Out of the box, it emulated a generic non-profit funder: competent, reasonable, and not remotely what I wanted. I realized that I don’t want generic help; I want help applying my own methods.
A funding style is a pattern of judgment. What types of leaders do you look for? When do you trust measurement, and when do you distrust it? When is a messy thing actually promising, and when is it just messy?
Few funders write down their full pattern. But most funders have parts of it written in pieces like memos, strategy documents, emails, etc. Your paper trail is your judgment’s fossil record.
I had a lot of fossils, over 150 pages of published and draft essays. I gave them to Claude and asked it to distill my funding "playbook" — the principles and questions I actually apply. It did a surprisingly good job, distilling principles such as:
Judge whether a project builds future movement capacity, not only its immediate cost per vote (Searching Beyond CPV).
Force projections into explicit Fermi estimates (The Impact Funnel).
Evaluate distribution as seriously as program measurement (Products vs Programs).
The AI started out hyper-attuned to everything I’ve ever written. I thought that was what I wanted, but I was wrong.
A playbook inferred from my writing reflects the written version of me, and the written version of me is incomplete. There are things I never wrote down. There are things I wrote with nuance that was obvious to me but invisible on the page. There are things I’ve stopped believing.
So the playbook is only a first draft of my judgment, reconstructed from its fossil record. Improving that draft is where it gets even cooler.
“I know kung fu”
Before AI, the only way to improve my funding judgment was to improve myself. I’d hear a useful idea, think “that’s important,” and hope I remembered it six months later at the exact moment it mattered.
Someone taught me a concept I’d never written down: evaluate the marginal dollar, not the average dollar. (A pitch quotes you the program’s average impact — but the first dollars in may have already bought the best sites and the easiest turf, and the next dollar buys a less-good increment.)
The old way: hope I remember to ask about the marginal dollar next time.
The new way: I add it to the playbook, and the question runs in every future review.
The strange part is that my funding process can now learn something I haven’t fully learned. I can encounter a useful distinction once, install it, and benefit from it later without trusting myself to remember it.
This reminds me of The Matrix, when Neo gets a program loaded into his brain, opens his eyes, and says, “I know kung fu.” This isn’t quite that; I won’t always remember kung fu, but my funding process will.
Making your own playbook
AI can infer your first-draft playbook from sources like:
Memos about your past funding decisions
Funding strategy documents
Emails explaining why you didn’t fund something
Ask what principles it sees, what questions you keep asking, where your reasoning seems underspecified.
If you don’t have enough written materials, you’re not stuck. Have it interview you, one question at a time: What kinds of organizations do you like to fund, and why? What are you more or less confident about? What do you distrust in a pitch? What have you funded that worked, or not? What do other funders over- and under-value? If you don’t like the results, tell it why and iterate.1
Conclusion
I started by trying to get better answers from AI. What I ended up with was a draft of how I make decisions.
That draft is imperfect. But unlike the version in my head, it can be inspected, criticized, compared with other versions, and deliberately updated.
The obvious AI use case in grantmaking is reading proposals. The more interesting one is improving the judgment that reads them.
PS for organizations
You have a new fundraising challenge. My playbook works poorly without enough information on an organization, because gaps get filled by whatever the AI can find on the web or elsewhere.2 I suggest developing a corpus of your materials that’s easily read by funders’ AIs. More on this to come; reach out if you want to discuss.
Thank you to Jonathan Robinson for giving me the point about average vs. marginal dollars for my playbook.
I suggest doing this verbally with a modern dictation tool such as Wispr Flow.
I tested this on an organization I'm close with: I gave Claude Opus the organization’s funder-facing materials plus web search, and asked it 23 funder-style diligence questions. On about 40% of the questions, Claude’s answer was materially worse than the organization’s real story: not false, but incomplete or unfair in ways better inputs can address.