How CRAI Built an Open-Source AI Grant Writing Tool for Nonprofits

An Interview with Jason Strickland, Head of Technology

The conversation that started CRAI's open-source AI grant writing tool for nonprofits happened over coffee. Jason Strickland had recently stepped away from a corporate job, looking to plug into something with more community purpose. A connection through mutual contacts brought him to CRAI, and a specific problem was already waiting: could he build something to help with grant research?

Who is CRAI?

The Center for Rural AI is a nonprofit working to turn rural communities into AI knowledge hubs. The work is practical: research, pilot programs, trainings, and tools designed to meet organizations where they actually are. For the full picture of what that looks like day to day, visit what we do.

The organization runs on a small team. Like most nonprofits its size, it relies on a handful of people doing a lot of different things. Jackson, CRAI's grant writer, is one of them. Grant writing for a nonprofit this size is rarely a focused, uninterrupted activity. It happens in between everything else, on deadlines, in stolen hours. The research phase alone, before a single word of the actual proposal gets written, can take days or weeks: finding opportunities, checking eligibility, assessing organizational fit, reading program guidelines. Jackson was skilled at the work. The research phase was also the problem: too many hours, before the actual writing even started.

That was the problem Jason was asked to solve.

Building an open-source AI grant writing tool without code

Jason's approach before building anything is to understand the workflow manually. He sat with the process, understood where the time was actually going, and started designing backward from what a useful output would look like.

The result is a five-skill grant writing assistant with no code. Every skill is written in plain language that a non-technical person can read, inspect, and change. The five skills handle setup, prospect scanning, fit scoring, data logging to Airtable, and audit logging. A staff member runs them in sequence. “You just hit go and then you just go back to your work.” In roughly 10 minutes, the tool returns a prioritized, scored list of prospects ready for the grant writer to review.

At the center of the tool is a 44-question organizational template. It captures who the organization is, what they do, who they serve, and how they describe their mission. Once filled out, Claude works from that template alongside whatever grant-related documents the organization already keeps in Google Drive to assess how well a funding opportunity actually fits.

AI as the installer

One of the cleaner design decisions in the tool is how setup works. Jason built a skill specifically so that AI handles the installation. Once an organization's accounts are connected to Claude, running the skill does the rest: it links the tools, configures the data, and establishes the workflow automatically. There is no manual setup. Plug in the dependencies, run the skill, and the system is ready, the same way every time.

Built to last, regardless of the model

Every design decision in this tool carries a version of the same question: if Claude gets replaced by something else tomorrow, does the tool still hold? The architecture has to say yes. Jason treats AI models as interchangeable. Swap one out, hand the new one the same memory and the same structure, and the work continues.

That philosophy shows up in the audit log. Every action the tool takes gets recorded, and that record belongs to the institution, not to a service someone else controls. If an organization switches models, their history travels with them.

“I've run it cold in Opus. I've run it cold in Fable. I've probably done half a dozen test installations on it and it's impressively been consistent every single time.”

Complexity is earned

Jason's working principle for the tool is direct: “Complexity is earned.” The system is as simple as it can be and no simpler.

The write skill uses hardcoded 16-digit Airtable field IDs rather than field names. That sounds fussy until you understand the reason: there is no ambiguity about where data lands. The prospect scanning skill runs on Claude's Fable model, which handles the creative work of finding overlooked funding opportunities. The install runs on Sonnet, because Sonnet follows instructions precisely. Different jobs, different tools.

Two constraints are hardcoded throughout: the tool never assumes, and it never guesses. “It's forbidden from assuming anything. If it can't find concrete evidence to answer the question, it will refuse to answer the question.” For grant research, where a poorly assessed fit wastes real time and real effort, that discipline matters more than appearing confident.

Try it and let us know

The grant writing assistant is open-source and available now at github.com/RuralAI/grant-assistant. If you download it and it starts making a dent in your grant research time, Jason has one ask: tell them. Comments, follows, and real feedback go a long way for a small team putting tools out into the world.

And if you're really feeling it, consider donating. CRAI is a nonprofit and the work continues because people choose to support it.

To learn more about CRAI's mission and the broader effort to build AI capacity in rural communities, the about page is a good place to start.