Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for Tax Season Receipt Pack
Inspiration
Every year, freelancers and families lose an evening or a whole weekend to the same chore: digging through download folders, scanner output and phone photos to find last year's receipts, retyping amounts into a spreadsheet, and hoping nothing was missed or sent twice. I live this as a freelancer in Argentina, where the accountant asks for "everything from last year" and the answer is spread across PDFs, scans and pictures of crumpled paper tickets. The documents are already on the NAS. The work of finding, reading and sorting them should be too.
What it does
You say: "Prepare my 2025 tax pack from the Invoices and Scans folders, and include receipt photos." The Skill:
- Confirms the year, the folders, whether to search the photo library, and your categories.
- Finds every candidate document and image in those folders, and photos of paper receipts in the photo library.
- Extracts date, vendor, document number, total, tax and currency from each one, and flags suspected duplicates.
- Runs OCR on images that could not be read, marking those values for review.
- Shows one table with totals per category and per month, per currency, plus a review queue: duplicates, unreadable files, receipts from another year.
- After you confirm, saves a Markdown summary and copies the included originals into an accountant-ready folder with one subfolder per category. Optionally tags the originals, for example tax-2025.
Originals are never moved, renamed, edited or deleted. Missing values are never guessed. Currencies are never mixed. It does not give tax advice.
How we built it
With the official Uliya Skill Kit, following its workflow: feasibility check against the capability catalog (version 2.2.0), init, generation, and the kit's static checker. The Skill is pure orchestration: the parent only delegates, to the single expert the catalog assigns to each CLI.
- file-agent: ug-filesystem find, file-workflow batch-transfer (copy only), ug-filemgr tags
- photo-agent: ug-photo search-screen and vision-ocr
- document-agent: ug-document extract with receipt, invoice and expense schemas and duplicate detection
- knowledge-agent: ug-knowledge save-text
Steps 1 to 4 are read-only. Every write needs a confirmation in the conversation and the platform's own workflow or gateway approval. Every delegation state (completed, failed, cancelled, waiting_user, waiting_approval), empty results and partial results are handled explicitly.
Challenges we ran into
- Designing for honesty: receipt file dates rarely match receipt dates, so the Skill filters by the extracted issue date and routes undated or out-of-year documents to review instead of guessing.
- Photo search can return previews without real file paths; the Skill reports those as found but not processed instead of inventing paths.
- Keeping the archive safe: copy only, no overwrite, no automatic retries after a cancelled copy.
Accomplishments that we're proud of
- One conversation replaces a weekend of manual work, while the user stays in control at a single review gate.
- Six NAS capabilities across four experts and four domains, combined into one everyday workflow.
- Static check result: static-pass with 0 errors and 0 warnings.
What we learned
A good NAS Skill is more about boundaries than steps: what to never touch, when to stop, and how to say "nothing found" plainly. The capability catalog made it possible to design every step against real, published operations instead of assumptions.
What's next
- Validate on a real UGREEN AI NAS with the synthetic test set in the package (two core tests and one edge case).
- Tune category suggestions from user corrections within a run.
- Companion Skills on the same pattern: a warranty and manuals register, and a home insurance inventory.
Built With
- claude
- file-workflow
- markdown
- ug-document
- ug-filemgr
- ug-filesystem
- ug-knowledge
- ug-photo
- ugreen-ai-nas
- uliya-skill-kit
