warranty-claim-pack

• 

  Snap the label, get the claim: label serial → invoice and receipt → warranty terms → claim pack

Inspiration

When an appliance breaks, the hard part of a warranty claim is not the claim: it is finding the proof. The invoice is a PDF buried in Downloads, the receipt is a phone photo of a paper ticket lost among thousands of pictures, and the warranty terms are on page 3 of a scanned manual. On a UGREEN AI NAS all of that already lives in one private place: the document library and the phone photo backup. We wanted the NAS to connect those pieces for you.

What it does

Snap the label, get the claim. You photograph the rating label of the broken product (or just name it) and warranty-claim-pack:

1. reads the brand, model and serial number from the label photo;
2. searches document content, file names and the text recognised inside photographed receipts, with multilingual purchase terms (invoice, factura, Rechnung, scontrino...);
3. links the matches into one evidence chain: label serial, invoice line, receipt photo, warranty terms, with a cited source for every fact;
4. reports the status, for example "Covered until 2027-03-14 · ElectroHome_order_INV-0425.pdf p.1, same serial as the label · 3-year term: manual p.3";
5. after your approval, creates a claim folder with copies of the evidence, a claim summary and a draft claim letter.

A read-only warranty check-up lists the warranties that end soon, and an optional "Proof of purchase" tag turns the NAS into a growing household warranty registry.

How we built it

It is a pure-Markdown orchestration Skill built with the UGREEN Skill Developer Package (Uliya Skill Kit 0.7.3): a SKILL.md plus five references, 15 delegate commands, and CLI ownership declared from the capability catalog 2.0.0. Four registered experts do the work: photo-agent (photo text search, vision model), file-agent (folders, file search, copies, tags), document-agent (invoice and receipt extraction, Q&A, letter draft) and knowledge-agent (content search, saving the summary). Discovery is read-only and runs in parallel; downstream steps only use real paths returned upstream; every write needs user confirmation and platform approval, never moves originals and never overwrites. All five delegation states (completed, failed, cancelled, waiting_user, waiting_approval) are handled. The package passes the official static checker with 0 errors and 0 warnings.

Challenges

• Matching the exact capabilities in the catalog: one search per keyword, no guessed paths, copies that can never overwrite, and an honest approval count for the user.
• Never guessing: ambiguous dates such as 03/04/2024, two purchases of the same model, or a missing warranty period all lead to a question, not an assumption. The Skill gives no legal advice.

What we learned

The value of an AI NAS is not one more search box, but joining evidence that already sits in different places: photos, documents and manuals, without anything leaving the user's device.

What's next

Validating on a real iDX6011 Pro with a synthetic test set (12 files, fictional brands) and the expected results in references/test-plan.md, then publishing to Skills Hub.

Built with Claude (Anthropic) via Claude Code for research, drafting and review, as disclosed in the package.

Built With

• artificial-intelligence
• claude
• markdown
• ugreen