Shoebox Tax Packet

Finds the receipts and invoices already on your NAS, including phone photos, and shows which expenses lack proof and which invoices have no payment on file.

Comment

What it does

Shoebox Tax Packet is a Skill for UGREEN AI NAS. It turns the receipts, invoices and receipt photos scattered across the NAS into a quarterly tax packet and a ledger that flag missing proof and unpaid invoices.

The problem

Freelancers who file taxes every quarter keep receipts in Downloads, email attachments and phone photos. They count a receipt twice when it exists as a PDF and a photo, and they miss unpaid invoices.

How it works

One request ("get my Q3 tax stuff together") runs four phases:

  1. Read (read-only): file-agent find, scan-dedupe and cat; photo-agent health, search-picture by text, vision-ocr and vision-vlm; document-agent extract, plus reconcile for uncertain pairs.
  2. Plan: the user reviews a plan, and the Skill writes nothing before the user says yes. From my simulation:
Income: 9 invoices, USD 17,250.00. Expenses: 31, USD 8,206.75 and EUR 41.65.
I merged 5 duplicates. Review: a locked PDF, a blurry photo.
Missing proof: 2 client invoices, 2 contractor bills. Copy 42 files?
  1. Write: after the user's yes and NAS approval, file-agent mkdir and cpmv copy the files into the packet, and knowledge-agent save-text saves the ledger. Originals stay where they are.
  2. Ask: later questions such as "which invoices are unpaid?" go to document-agent qa on the ledger.

Required inputs: the quarter and the business names. Reruns copy only new files and reuse earlier decisions. The experts mask card numbers and tax IDs, and the Skill copies nothing off the NAS.

What is different

Receipt apps make you capture each receipt into their cloud. The NAS already backs up your receipt photos, and Photos reads their text. Paperless-ngx archives files but finds no gaps. This Skill builds the gap list an accountant asks for, links every amount to its source file, and reuses your decisions.

How I built it

The Skill is Markdown only: SKILL.md plus four references (packet and ledger rules, setup and removal, test scenarios, disclosures). I made it with the official UGOS Skill Creation Tool (uliya-skill-kit 0.7.3), and the kit's static checker reports no errors. In a laptop simulation, Claude (sonnet), standing in for Uliya with scripted experts, passed 119 of 119 checks in 10 scenarios.

AI tools: I used Claude (Anthropic) for writing help.

Status and limits

The Skill has not run on a UGREEN AI NAS yet. I have not tested approval prompts, expert output or crumpled receipts on a NAS. It needs UGOS Skills support and Photos text recognition; without it, it reads only documents and folder images. Accuracy depends on NAS OCR, and the NAS may ask to approve each copy. The Skill does no bank matching or currency conversion, reads up to 500 documents and 300 photos per quarter, and gives no tax advice.

What's next

If selected as a finalist, I will:

  • build a packet from 30 receipts I photograph plus test PDFs, then check fields, merges, ledger links and original checksums;
  • add two receipts and rerun; the Skill must copy only those and reuse my decisions;
  • check that a blurry photo and a locked PDF go to Needs review without guessing.

Built With

  • markdown
  • ugos-skill-creation-tool
  • uliya-skill-kit
Share this project:

Updates

Submission history