Inspiration
In a savings circle, a group of friends each pay in the same amount every cycle and one member takes home the whole pot. It is called a chama in Kenya, a tanda in Mexico and a susu in West Africa and the Caribbean. I live in Kenya, where chamas are everywhere.
Every circle has a volunteer treasurer. Each cycle they collect proof of payment in whatever form members send it: statements, receipts, screenshots of payment confirmations. Then they check it all against the member list by hand. It is slow, easy to get wrong, and awkward between friends.
What it does
The treasurer says something like "reconcile March contributions for Sunday Savers". The Skill then:
- finds the member roster and that period's payment evidence in the folder the treasurer names on the NAS;
- extracts each payment from documents and reads payment screenshots with OCR;
- matches payments to members using fixed rules, and never guesses a match;
- shows four lists: paid in full, short or over, no payment found, and payments it could not match;
- saves a cycle report to the same folder once the treasurer approves, with a source file cited on every line.
It only reads until the treasurer approves the save. It never moves or deletes evidence, never contacts members and never moves money.
How we built it
It is a Markdown-only orchestration Skill written to the Uliya Skill Development Standard with UGREEN's Skill Developer Package. It delegates to four platform experts: file-agent to find files and apply tags, document-agent to extract payments, photo-agent for OCR on screenshots, and knowledge-agent to save the report.
The package passes the kit's static checker. It has not yet run on a NAS; that is what Round 2 is for.
Challenges we ran into
- Deciding what the Skill should refuse to do. A member with no matching payment is reported as "no payment found in these files", never as "has not paid".
- Handling doubt: low-confidence amounts, duplicate receipts, and two members who share a first name.
What we learned
The useful part of the design is mostly the rules for when to stop and ask, not the extraction itself.
What's next
Testing on a real UGREEN AI NAS with synthetic circle data, especially OCR accuracy on payment screenshots.
AI tools used
Claude (Anthropic) was used to read the capability catalog, draft the Skill files and run the kit's static checker.
Built With
- claude
- markdown
- ugos
- uliya