What a nonprofit brain is, and how to set one up

A chat window with no context produces generic prose. A chat window that has read your mission, your outcomes data, your last six proposals and your style guide produces a first draft you can edit. The model is the same in both cases. The difference is memory, and memory is something you build.

That is what this course means by a nonprofit brain: a single, maintained place where your organization’s knowledge lives, with instructions that tell an AI assistant how to use it, and a small set of repeatable tools that run on top of it. By the end of this course you will have one, built in Claude, plus three working tools: a prospect pipeline, a proposal drafter, and a reporting workflow.

Chatting versus a brain

Most people’s first month with an AI assistant looks like this: open a new chat, paste in the mission statement, paste in the funder’s guidelines, ask for a draft, get something that sounds like every other nonprofit, fix it by hand, close the tab. Next week, repeat from zero.

Three things are wrong with that loop:

1. Nothing accumulates. Every chat starts empty. The fix you made to how the draft describes your service area is gone.
2. The facts are whatever you remembered to paste. If you forgot the outcomes table, the model fills the gap with something plausible. Plausible is the problem.
3. The voice is the model’s default. Without examples of how your organization writes, you get the average of the internet.

A brain fixes each one with a specific mechanism:

Keep that table. The rest of the course fills in each row.

The three layers

Think of the brain in three layers. Build them in this order.

Layer 1 — Memory (lessons 2 and 3). Ten documents that describe your organization accurately: fact sheet, mission, theory of change, programs, outcomes, budget narrative, boilerplate, past proposals, funder notes, voice guide. These go into a Claude Project as project knowledge.

Layer 2 — Instructions (lessons 1 and 3). A page of standing rules the assistant follows in every chat in the Project: who it is working for, which documents are authoritative, how to handle a number it cannot find, what it must never do.

Layer 3 — Tools (lessons 4 to 6). Repeatable procedures: find and triage prospects, draft an LOI or proposal section, turn program data into a funder report, board memo, or thank-you letter. You will build these as a prompt library first and, where it helps, as Claude Skills.

Guardrails (lesson 7) wrap all three. The capstone (lesson 8) runs one real proposal through the whole thing.

What Claude calls these things

Product names change. As of September 2026, this is how Anthropic’s help center describes the features this course uses:

• Projects. A workspace with its own chat history, a knowledge base you upload documents into, and custom instructions that apply to every chat inside it. Anthropic notes that context “is not shared across chats within a project unless the information is added into the project knowledge base.” That sentence is the whole argument for building a brain: if it matters, it goes into knowledge.
• Project knowledge and RAG. On paid plans, when project knowledge approaches the model’s context limit, Claude switches to retrieval mode automatically, which Anthropic says expands capacity by up to ten times. Practically, you can hold a full development office’s worth of documents in one Project on a paid plan.
• Project instructions. Set with “Set project instructions” inside a Project. They apply to every chat in that Project.
• Skills. Folders of instructions (a SKILL.md file plus optional resources) that Claude loads when a task matches the skill’s description. Skills require the “Code execution and file creation” capability to be on.
• Connectors. Links from Claude to other services such as Google Drive, Microsoft 365 or a CRM, and to “custom connectors” that point at a remote MCP server. You switch them on per conversation.

Choose a plan, deliberately

Your plan decides three things that matter here: how much knowledge a Project can hold, how many custom connectors you can add, and whether your conversations can be used for model training.

Sources for each cell are at the bottom of the lesson. Check them before you buy: plan names and limits change.

Decision rule. If you are one person testing whether this is worth it, the Free plan can hold a small brain, but you will hit the five-project and knowledge limits quickly. If more than one person will touch grant work, or anything in your documents is internal, use a Team plan so the commercial terms apply and colleagues can share the Project.

Claude for Nonprofits

In December 2025 Anthropic launched Claude for Nonprofits. As of September 2026 its pages state:

• Discount. Anthropic’s announcement describes discounts of up to 75% on Team and Enterprise plans. Its nonprofit page lists a Team “Nonprofit Standard” price of $8 per user per month with a two-seat minimum, and a lower “Nonprofit Access” price for organizations in low- and middle-income countries. Enterprise pricing is through Anthropic’s sales team.
• Who is covered. Anthropic lists registered 501(c)(3) organizations and international equivalents, K–12 schools, and certain mission-based healthcare providers. Its tutorial lists government agencies, political campaigns, higher education and healthcare systems as not covered. Anthropic makes that determination, not this course.
• Verification. Team customers verify through Goodstack, which Anthropic says takes a few minutes. Enterprise customers are verified by Anthropic’s sales team.
• Connectors. Nonprofit-specific connectors for Blackbaud (Raiser’s Edge NXT), Candid (nonprofit and funder data) and Benevity, alongside general ones such as Google Workspace, Microsoft 365 and Slack.
• Training. A no-cost course, AI Fluency for Nonprofits, built with GivingTuesday.

Treat the fluency course as the prerequisite for this one. It teaches how to work with AI well. This course assumes that and teaches the build.

Set it up: the procedure

Do these in order. Budget about 45 minutes.

1. Pick the account owner. Use an organizational email, not a personal one. If the ED leaves, the brain should not leave with them.
2. Choose the plan using the table above. If you are going the Team route, start at claude.com/solutions/nonprofits and complete verification.
3. Set privacy before you upload anything.
   • On an individual plan (Free, Pro, Max), open Settings, then Privacy, and decide on the model-improvement setting. Anthropic’s consumer terms let you choose whether chats are used to improve models; opting in extends retention to five years for those chats, and opting out keeps the 30-day period. For an organizational brain, opt out.
   • On a Team plan, Anthropic states it does not use Team data for training by default. Confirm your organization settings anyway.
4. Turn on the capabilities you will use. Settings, then Capabilities: turn on “Code execution and file creation” (needed for Skills and for Claude to produce Word, Excel and PDF files). On Team, an owner enables this and Skills under Organization settings.
5. Create the Project. Go to Projects, choose ”+ New Project”, name it for the organization — for example Riverbend Brain — and write a one-line description: “Riverbend Food Collective’s knowledge base for grants, reports and donor communications.” On Team, choose “Keep it private” until the knowledge is clean, then share.
6. Paste the starter instructions below into “Set project instructions”. You will replace them with the full template in lesson 3.
7. Upload nothing else yet. Lesson 2 is about which documents go in and in what shape. Uploading your whole shared drive now is the most common way to get a brain that confidently cites a 2019 budget.

Starter project instructions

Copy this exactly, swapping in your organization’s name:

You are the development assistant for [Organization Name], a nonprofit.
You work for our staff. Your job is to draft, summarize and check
grant, report and donor materials using ONLY the documents in this
project's knowledge base and what the user tells you in the chat.

Rules:
1. Never invent a number, date, name, partner, quote, or outcome.
   If a fact is not in the knowledge base or the chat, write
   [NEED: what is missing] and keep going.
2. After any sentence containing a number, add the source document
   name in brackets, e.g. [src: 05-outcomes-data].
3. Never state or imply that we are eligible for, or likely to win,
   any funding. Quote the funder's published criteria instead.
4. Everything you produce is a draft for a named staff member to
   review. Say so at the top of every draft.
5. If a request involves personal information about a client,
   beneficiary or donor, stop and ask whether it should be here.

One brain or several?

A Project cannot read another Project’s knowledge. That shapes the architecture:

• A single organization: one brain Project. Resist making a Project per funder or per program; you would split your facts across places that cannot see each other. Use chats inside the one Project for each piece of work.
• A grant consultant: one brain Project per client. Client facts must never mix. Your own tools (the Skills in lessons 5 and 6) are account-level, so they work in every client Project without being copied.
• A sensitive side-workspace: if the board wants a separate Project for, say, executive-session material, fine, but keep it out of the grant brain entirely.

What this course does not do

It does not make AI your grant writer. Every tool you build produces a draft that a named person reads, corrects and signs off. It does not assess eligibility for any grant; the funder decides that, and lesson 4 shows how to read their criteria. And it does not replace relationships with program officers, which remain the part of fundraising no model touches.

Do this now in OpenGrants

Lesson 4 builds your prospect pipeline in OpenGrants, and the capstone needs it. Start now so your saved searches have a few days to accumulate.

1. Create an account at ops.opengrants.io/signup. Every new account starts with a 7-day free trial; after that the app is $9/month.
2. If you are a 501(c)(3), email your EIN to [email protected] to request the verified nonprofit rate of $4.50/month (50% off). Support confirms verification.
3. Fill in your organization profile using the same wording you will put in your fact sheet in lesson 2.

In Claude: create the Project, set privacy, and paste the starter instructions.

Start your OpenGrants trial (opens in a new tab)

Check yourself

1. Name the three layers of a nonprofit brain and the order to build them in.
2. Why does a fact mentioned in last week’s chat not help this week’s draft, and where should that fact go instead?
3. On which plan types are your chats not used for training by default, and where do you set this on an individual plan?
4. A consultant with eight clients asks whether to build one Project or eight. What do you tell them, and why?
5. Which capability must be on before Skills work?

Sources

• Claude for Nonprofits — Anthropic — December 2025 launch, discounts of up to 75% on Team and Enterprise, Blackbaud/Candid/Benevity connectors, AI Fluency for Nonprofits with GivingTuesday
• Nonprofits — Claude — eligible organization types, Goodstack verification, Team nonprofit pricing and seat minimum, connector list
• Getting started with Claude for Nonprofits — Claude Academy — exclusions, Team/Enterprise training default, connector descriptions
• AI Fluency for Nonprofits — Anthropic Academy — the course itself
• What are projects? — Claude Help Center — Projects on all plans, five-project Free limit, retrieval mode on paid plans
• How can I create and manage projects? — Claude Help Center — creation steps, knowledge not shared across chats, instructions, Team sharing
• Use skills in Claude — Claude Help Center — Skills availability and code-execution requirement
• Get started with custom connectors using remote MCP — Claude Help Center — custom connectors by plan, one on Free
• Updates to Consumer Terms and Privacy Policy — Anthropic — training choice for Free/Pro/Max, retention periods, commercial plans excluded