Inspiration

In Korea, a rental deposit is often many times the monthly rent, and for students it can be most of what they own. At move-out, landlords sometimes deduct for marks that were already there on day one. Most tenants did take photos when they moved in, but two years later those photos are lost among thousands of phone backups, nobody can prove when they were taken, and nobody remembers what the lease said about cleaning or wear and tear.

A NAS already holds both halves of the answer: the lease PDF and the automatic phone photo backup. Deposit Shield connects them.

What it does

Say "I'm moving out, prepare my deposit evidence." Deposit Shield then:

  1. Finds the lease on the NAS and lets you pick the right file.
  2. Reads the deposit terms (amount, dates, return deadline, cleaning, repairs, wear and tear, deductions), each with the page number and the exact sentence.
  3. Finds the photos and videos taken in the move-in window (and the move-out window, if you have one) by capture date.
  4. Checks each photo's original EXIF capture time and records a SHA-256 fingerprint, so you can later show the file was not edited.
  5. Writes a short, neutral note on what is visible in each photo, without blaming anyone.
  6. After your approval, creates a private album with exactly those photos.
  7. After your approval, saves a Markdown report you can show the landlord.

Photos without EXIF capture times are kept but listed separately, so you never present them as dated proof. Missing lease fields say "not found" instead of being guessed. The pack is evidence, not legal advice.

How we built it

Deposit Shield is an orchestration Skill that follows the Uliya Skill Development Standard. SKILL.md delegates seven ordered steps to four registered experts:

  • ug-filesystem → file-agent: find the lease, SHA-256 fingerprints
  • ug-document → document-agent: extract and qa for lease terms with citations, writing for the summary
  • ug-photo → photo-agent: search-picture by capture date, meta for EXIF, vision-vlm for condition notes, create-album
  • ug-knowledge → knowledge-agent: save-text for the report

Each step uses only the real result of the step before it, and all five delegation states (completed, failed, cancelled, waiting_user, waiting_approval) are handled. The package passes the official UGOS Skill Creation Tool checker with zero errors and zero warnings.

Challenges we ran into

Evidence has to be honest to be useful. The hardest part was writing the workflow so it never fills a gap: no widening the photo search window without asking, no estimated amounts, no "damage" judgments, and a clear line between EXIF-dated and undated photos.

Accomplishments that we're proud of

The whole flow is read-only until the two moments where it changes something (the album and the report), and both go through platform approval.

What we learned

How UGREEN's expert-delegation model separates "what the user wants" from "how the NAS does it", and how much a NAS's photo index and document reader can do together.

What's next for Deposit Shield

If selected as a Finalist: test on a real UGREEN AI NAS with a synthetic lease and test-room photos (move-in pack, move-out comparison, and an edge case with a missing deadline and EXIF-stripped photos), then tune the default date windows and the condition-note prompt.

Disclosures

No third-party components; the package contains only Markdown. AI tools used: Claude Code (Anthropic, Claude Opus 5.5) for suggesting the idea, reading the starter kit and catalog, drafting SKILL.md, its references, and the proposal, and running the official checker. Full disclosures are in the Skill package (references/disclosures.md) and the proposal PDF.

Built With

  • claude-code
  • markdown
  • ugos
  • ugreen-ai-nas
  • uliya-skill-kit
Share this project:

Updates

Submission history