How to build an AI grant team with Claude

Most people use AI by prompting: ask a question, watch it spin, copy the answer out. That keeps you in the chair. An employee is different. It knows its role, it holds standing context about your organization, it follows your rules, and it files its own work. You hand it the job once, in one session, and it runs the whole sequence.

Concretely: my grant writer creates the file path itself. Grants folder, then 2026, then the funder's name, then the opportunity. Into that folder it puts a draft, a budget, a checklist, and notes. I did not do any of that. I gave it a job.

This is the hardest build, and it is first because everything downstream depends on it. Its job is to find opportunities, verify they are real and open, score them against your reality, and put them into a pipeline you can act on.

What you give it

• Role and context: who it is, what your organization does, who you serve, where you operate.
• Your annual budget, so award size can be judged relative to something true.
• Search rules: prioritize funders with a history in your issue area, skip funders who only give to universities or hospitals.
• A verification rule: confirm every deadline on the funder's own website, and never return an opportunity you cannot confirm is open.
• Fit criteria: mission alignment, award size relative to budget, clear eligibility, and how heavy the application is.

What it gives back

Twelve candidates cut to a strong ten, each scored one to ten with a sentence of reasoning and any deal breakers flagged. Then a spreadsheet with funder, program, award range, deadline, fit score, deal breakers, and a specific next action, sorted by fit. Then a short memo telling you what the sheet reveals that a list does not: which opportunities need a phone call instead of an application, which deadlines fall inside forty five days, and which award ranges are unpublished so you should not forecast revenue from them yet.

I have paid for Instrumentl. I have paid for GrantStation. I do not pay for either anymore, because this is how I run research and build pipelines now.

The writer takes a specific opportunity and produces a package. Feed it the blank application questions copied out of the funder's portal, the RFP, and the funder's budget template. Give it standing access to your approved organizational materials, and approved means your organization knowingly agreed to those documents going into AI.

Then it returns four things in the right folder:

• A full proposal draft written to the actual questions.
• A budget recoded from your operating budget into the funder's template, inside every constraint, including nested indirect cost caps.
• A submission checklist.
• Notes on what it still needs from you.

Then you edit with a fine tooth comb. That is not a caveat, that is the design. The draft is entry level work product. Your judgment is what makes it fundable.

The manager owns the state of your grant program: what is open, what is drafted, what is submitted, what is awarded, what reports are due and when. It is the employee that keeps the fall grant drop from becoming a crisis, because it always knows where everything stands.

If you have never had that, here is the honest test. Right now, can you say without opening anything how many proposals are in draft, what is due in the next thirty days, and which awarded grant has a report due next? The manager is the answer to that question, every morning, without you assembling it.

This is the step people skip, and it is the one that produces the real change. When the grant writer finishes a proposal, it goes into the grant pipeline and updates it to say proposal drafted. The prospector's verified opportunities flow into the writer's queue. The manager reads from both.

Once they share context, you stop starting over every time you open a session. They already know what has been done, what is completed, and where things stand today. It is close to sitting down for a staff meeting with them every morning, except the meeting takes two minutes and nobody is behind.

• Foundations first. Learn how the environment works, get your organizational context loaded, and confirm your setup before you build anything.
• Then the prospector and analyst. The toughest build. Everything downstream needs its output.
• Then the writer. Faster than you expect, because the context is already in place.
• Then the manager, and the connections. Wire them together last, once each one works on its own.

I taught all of this to myself, and it took six to nine months of trial and error to get to the level I operate at now. It does not have to take you that long. In the first cohort, a student shipped a working employee and submitted a grant she had already decided to skip this year, before the final session.

• A paid Claude plan, around $20 a month. The work happens in the desktop app on your own computer.
• If IT manages your machine, get approval before you install anything.
• Organizational approval to put budgets and documents into AI.
• Your core materials gathered: annual budget, program descriptions, past proposals, determination letter.
• A clear line for yourself on what stays human: your judgment, your relationships, your program truth.

Five to ten hours a week back is the point, and what you do with it is the real outcome. Get back into your community doing impact driven work. Or go for a run, work on the garden, call the friend you have not called. Life is meant to be lived, not worked, and the nonprofit sector is very good at forgetting that for the best possible reasons.