Lessons from Building Changefinder (AI for Grant Writing)
I built Changefinder, an AI-powered tool to help people write better grant applications. Users could paste in a grant question and their draft response, then get tailored feedback on clarity, alignment, and a suggested revision.
✅ What Worked
People actually won money
Users reported over $850,000 raised using drafts improved through Changefinder.
Simple, useful output
The combo of a score, specific feedback, and an actual rewritten suggestion made it far more actionable than a typical AI summary.
Letting users add context
You could specify the type of grant (education, startup, equity, etc.) or upload a rubric. That made the feedback much more specific and real.
❌ What Didn’t Work
Selling to foundations
The idea was that foundations would offer Changefinder to applicants to improve quality. But most were under-resourced and didn’t want to add anything to their workflow.
Training a model
I trained a model and it flopped. A strong system prompt with logic behind it worked just as well. Fine-tuning added cost and complexity without meaningful benefit.
UI was harder than expected
Even though it was just a wrapper around the prompt, a clean, simple interface made all the difference. Grant writers aren’t always technical, so the product had to feel dead simple. That took more iteration than I planned.
💡 What I Learned
UX matters, even for prompt wrappers
It’s not just about what the AI says. It’s how you frame it, display it, and guide people through it.
Grant competition is fierce
With the current funding landscape, applicants are desperate for help. But the process is still chaotic, and there’s a real gap between demand and scalable adoption.
Be ready to pivot fast
This landscape is changing constantly. You need to be able to switch up your underlying model, train a new one, or refactor code quickly to keep up with the "new best thing."
Talk to real grant writers
In order to build something that works, you need to talk to as many grant writers as possible. My boss is a great grant writer but was juggling a full-time job. That left me waiting on feedback, reaching out to people I didn’t know, or just guessing what they’d want. Don’t guess. Ask.
I’m open to sharing the system prompt or open-sourcing the code if people are interested. I’m also working on a rubric evaluator that scores answers using your own uploaded rubric shown below.
Happy to share more or chat if this is useful.
thanks to my wonderful team:
https://www.linkedin.com/in/ianwinbrock/
https://www.linkedin.com/in/madelinejtomka/
https://www.linkedin.com/in/phanisaipvv/
https://www.linkedin.com/in/harry-herzberg-ba000814b/ (me)
– Harry