Inspiration

After a break-in or a stolen parcel, insurers and police ask for the same three things: when it happened, proof you owned the item, and proof of what it cost. On a UGREEN AI NAS, all three usually already exist: camera events, years of photos, and scanned receipts. They're scattered across apps, though, and you have to dig through them on the worst day of your week.

So the idea was simple: one sentence in, one evidence-ready claim folder out.

What it does

"Someone took my parcel from the porch yesterday between 1 and 3 pm."

From that, Home Incident Claim Pack:

  1. Finds camera events in the incident window and lets you pick the relevant ones.
  2. Finds photos taken before the incident that show the affected items (proof of ownership).
  3. Finds receipts and invoices for those items and extracts merchant, date, amount, serial and warranty (proof of value).
  4. Shows one review list, including what's missing for each item, and waits for your approval.
  5. Copies the approved files into Incident-<date>-<type>/ with camera, photos and receipts subfolders, and saves an incident-summary.md ready to attach to a claim.

Originals are never moved or deleted, unknown people are never identified, and nothing leaves the NAS.

How I built it

It's a pure orchestration Skill built with the official uliya-skill-kit. Every step maps to a capability in the public NAS catalog:

Step Expert Capability
Camera events security-monitor-agent ug-surveillance event-timeline, event-evidence
Ownership photos photo-agent ug-photo search-picture, meta
Find receipts file-agent file-workflow get-user-dir, ug-filesystem find
Read receipts document-agent ug-document extract (receipt / invoice schemas)
Build folder file-agent ug-filesystem mkdir, cpmv (copy)
Write summary document-agent, knowledge-agent ug-document writing, ug-knowledge save-text

Each delegation is self-contained and passes only completed upstream results downstream. Every state change sits behind waiting_approval.

Challenges

  • Getting past the static checker. The section headings and capability-evidence format in the template don't match what the checker actually validates. I read the checker's rules directly and rewrote the Skill to pass with zero errors and zero warnings.
  • Partial evidence is the normal case. Real claims rarely have footage, a photo and a receipt for every item. The Skill treats any subset as a valid pack and reports what's missing per item instead of failing.
  • Safety boundaries. A claim tool should never move originals, guess values, or name strangers in footage, so those became hard rules in every delegation.

What I learned

Orchestration Skills live or die on the hand-offs. Being precise about what each expert returns, and what happens when it returns nothing, mattered more than the "AI" part.

What's next

  • Test object-search recall and receipt extraction on real hardware (Round 2).
  • Export camera snapshots into the pack once a supported export capability exists.
  • A "renter mode" checklist for landlords and deposit disputes.

AI disclosure: Claude (Anthropic) was used to help draft and check the Skill against the capability catalog.

Built With

  • claude
  • file-workflow
  • markdown
  • ug-document
  • ug-filesystem
  • ug-knowledge
  • ug-photo
  • ug-surveillance
  • ugos
  • ugreen-ai-nas
  • uliya-skill-kit
Share this project:

Updates

Submission history