Inspiration
Every month-end, freelancers and small businesses have to show a document for every payment on their bank statement. The receipts and invoices are usually already on the NAS, but checking them means opening dozens of PDFs and phone photos by hand, and the gaps surface late, when the accountant asks.
What it does
Paper Trail starts from the money that moved. You name one bank statement, up to five folders and a month. For each statement line it finds the document that proves it and returns a close packet:
- Matched: the line, the document, and the fields that decided the match
- Likely match: a candidate with the reason it is not certain
- Missing evidence: a payment with no document
- Evidence without payment: an invoice nobody paid
- Possible double payment and duplicate evidence
- Needs review: unreadable scans, two equal candidates, a currency that cannot be checked
It only reads. Originals are never moved, renamed or changed; the single write is the packet, saved once after you confirm and the NAS approves. Financial data never leaves the NAS.
How we built it
Paper Trail is a Markdown-only UGOS Skill. SKILL.md delegates eight steps to the NAS experts: file-agent (scope, find, identical-file scan), document-agent (statement and invoice extraction with the mixed_reconciliation schema, reconcile for doubtful pairs), photo-agent (OCR of receipt photos) and knowledge-agent (saving the packet, with workflow approval). Each step uses only the previous step's completed result. The matching rules are explicit: exact amount, a date window, plus a matching reference or counterparty; split, partial and double payments are defined. The package passes the UGOS Skill Creation Tool's static checker with zero errors and zero warnings.
Challenges
Writing rules that never guess: two equal candidates are shown, not chosen; foreign-currency payments are flagged instead of converted with an invented rate; unreadable photos go to review instead of being estimated.
What's next
A fictional test set (a design studio's statement and 20 invoices and receipt photos, three scenarios with expected results) is ready for validation on a real AI NAS.
Disclosure
Independent community Skill, not made or endorsed by UGREEN. AI tools: Claude Code (Anthropic), used to draft the Skill text, matching rules and test data and to run the checker. All examples and test data are fictional.
Built With
- claude-code
- markdown
- ugos
- ugreen
