Retail, 4 November 2026
Mid-Ohio Food Collective: Claude Code Cuts Feature Delivery From Months to Hours on a 15-Year-Old Codebase
Mid-Ohio Food Collective — Ohio’s largest food bank, serving a 20-county region through more than 500 partner agencies. Its software and data team built and maintains FreshTrak, a platform that manages pantry visits, eligibility checks and reporting — now licensed to 22 food banks across the United States, serving about 35,000 families a day.
Problem
Donated food is shrinking as grocery chains get better at forecasting their own supply, just as demand keeps climbing past pre-pandemic levels. FreshTrak, the software this runs on, is more than 15 years old, and the team’s development bandwidth never matched the backlog — ideas sat untouched for years, according to the team. Scaling FreshTrak to a new food bank meant hand-adapting it every time: each U.S. state has its own eligibility rules, reporting formats and consent language, so what worked in Ohio couldn’t just be copied to Texas.
Result
97%
- Feature delivery dropped from two to three months down to 24-36 hours — an idea reaching production roughly 97% faster.
- Onboarding and training time for a new partner food bank fell by half.
- A technical-support team member with no engineering background got her first pull request into production in under 12 hours; “She’s not an engineer,” said product lead Sarah Mollenkopf. “She is now.”
- A business analyst turned a single food-bank visit into a working prototype over one weekend.
- FreshTrak now runs in 22 food banks nationwide, logging more than 12.5 million household visits a year and 68 million since launch.
Solution
Engineers now brief Claude Tag directly in Slack; it works through the FreshTrak codebase on its own and opens a pull request. Claude Code takes on the bigger jobs — multi-service changes and migrations inside a codebase older than most of the people onboarding onto it. Nothing merges without an engineer’s approval, a passing test suite and at least one round of Claude-assisted review, and local staff still check every state’s eligibility rule against official agency guidance before it ships — the AI writes the code, people still own the rules. Lead engineer Nick Davis described the difference from earlier AI tools bluntly: Claude “understood the whole system” well enough to be “reviewed, challenged and sent back” like a senior engineer’s own work, rather than needing hand-holding through it.
Could a small team do this?
Mostly yes, and the honest reason is that the hard part here isn’t the AI — it’s the 15 years of accumulated business rules sitting on top of it, which no solo operator starts with. The actual technique transfers cleanly: pointing Claude Tag at a real codebase in Slack so it can propose a fix without anyone opening an IDE, and using Claude Code for the larger refactors, is exactly the setup a one-person shop can run today, on a codebase of any age, for the price of a Claude subscription. The “non-engineer’s first PR in 12 hours” result is the most buildable piece of all — it says the limiting factor is how well the person writes the request, not how much they already know how to code, which is the whole premise of a solo AI operator doing engineering work at all. What a solo team genuinely couldn’t replicate overnight is the multi-state compliance layer — the different eligibility rules, reporting formats and consent language for every jurisdiction — because that’s accumulated domain knowledge, not something Claude invents on day one; building that out would mean weeks of interviewing the actual rules before AI could help apply them. The transferable lesson for a Georgian solo operator is narrower and immediate: use Claude Tag in Slack (or its equivalent) as the first responder on your own code, and measure whether a non-technical teammate’s request can become a shipped change in under a day.