Inspiration
Claims get rejected for boring reasons. The receipt cannot be found. The serial number on the invoice does not match the one on the device. The insurance policy started a week after the incident. None of this is about whether the claim is honest. It is about paperwork.
People who own a NAS usually have the proof already: a scanned receipt, a warranty PDF, phone photos backed up automatically, recordings from a door camera. It is just spread across folders, the photo library and the camera history, and nobody checks whether the pieces agree until an insurer or a seller does it for them. I wanted the NAS to do the collecting and the checking before the claim is filed.
What it does
Claim Evidence Pack builds one evidence pack for one item and one incident. It supports five kinds of claim: theft, accidental damage, warranty fault, rental deposit dispute and delivery dispute.
The user says something like "build a claim pack for my stolen laptop, 3 March, search my Documents folder". The Skill then:
- Tells the user which sources it will read and asks for consent.
- Finds the documents for that item, including scans with meaningless file names.
- Reads receipts, invoices, warranties, policies and leases, and pulls out the serial number, dates and amounts.
- Quotes the clauses that decide the claim: coverage period, exclusions, reporting deadline and the proof required.
- Finds photos of the item from before and after the incident and reads any label or serial number visible in them.
- Looks up camera events around the time of the incident.
- Compares everything and gives each fact one label: confirmed (two sources agree), single source, conflict (sources disagree), missing, or not checked.
- Shows a draft, and only after the user approves saves a dossier, with optional copies of the original files and a photo album.
When two sources disagree, the pack shows both values and where each came from. It never picks a side. It also lists what is still missing, so the user knows what to look for before filing.
Everything runs on the NAS. No data is sent to outside services, no faces or number plates are looked up, and no existing file is changed or deleted.
How we built it
The Skill was built with the official UGOS Skill Creation Tool (uliya-skill-kit 0.7.7), following the Uliya Skill Development Standard. It is a SKILL.md file with three linked reference files: the cross-check rules, the dossier layout and the disclosures. There are no scripts or executables.
I started from the public capability catalog, which lists 9 CLIs and 84 actions, and designed the workflow only around capabilities that exist there. The Skill uses six CLIs (ug-filesystem, file-workflow, ug-document, ug-knowledge, ug-photo, ug-surveillance) through five registered experts (file-agent, document-agent, knowledge-agent, photo-agent, security-monitor-agent). The parent Skill only delegates. Six read steps run first, then the Skill cross-checks the results itself, and three write steps run last, each behind platform approval.
The package passes the official static check with zero errors and zero warnings.
I used Claude, through Claude Code, to read the capability catalog, suggest candidate ideas, draft the Skill files and the proposal, and check them against the tool's rules. This is disclosed in the package and in the proposal.
Challenges we ran into
- No scripts. A Skill cannot contain code, so all the comparison logic, such as how to treat a serial number that differs only by an easily misread character, had to be written as plain instructions.
- No NAS in Round 1. Nothing could be run on a device, so every step had to be checked against the catalog and written so that a missing or empty result stops the steps that depend on it.
- Receipts with useless names. A file called scan_0042.pdf will never match a search for "laptop", so the Skill falls back to a content search inside the chosen folder.
- Privacy. The Skill touches photos and camera history, so it asks for consent up front, limits itself to the folder, photos and time window the user chose, and never identifies people.
- Honest gaps. "Nothing was found" and "this step did not run" look the same in a careless report. The Skill keeps them apart as missing and not checked.
Accomplishments that we're proud of
- It checks whether different kinds of evidence agree with each other, which a chat summary of one file or a plain file search does not do.
- It reports conflicts instead of hiding them.
- It reads first and writes last, and every write waits for approval.
- It uses five NAS capability areas in one workflow: files, knowledge, documents, albums and surveillance.
What we learned
- How Uliya Skills work: the parent Skill delegates to registered experts and must handle five outcomes for each delegation: completed, failed, cancelled, waiting for the user and waiting for approval.
- A completed step is not the same as a useful result. The next step should only run if the exact path, value or reference it needs actually came back.
- A saved file is not automatically searchable in the knowledge index, so the Skill reads its own dossier by exact path.
What's next for Claim Evidence Pack
- Test on a real UGREEN AI NAS if selected as a Finalist: a theft claim where the serial number should come out confirmed, a warranty claim where the serials differ and should come out as a conflict, and a folder with no matching documents.
- Tune the wording of each delegation based on how the experts actually respond.
- Test the update path that adds new evidence to an earlier pack without replacing it.
- Add more claim types once the first five are proven on a device.
