Inspiration
Every deposit dispute comes down to one question: was it already there? In 2025, damage and cleaning made up 48% of deposit claims at the UK's Deposit Protection Service, and both are judged against the move-in condition. Most tenants did photograph move-in day. Years and phones later, though, those photos survive only in a NAS backup among tens of thousands of others. Finding the day is easy; matching each deduction line to a photo of that exact spot is not.
What it does
It Was Already There is a UGOS Skill for Uliya. The tenant names the landlord's letter and their move-in date. The Skill then:
- reads the letter into deduction lines;
- lists every photo from the move-in window, and looks at each one once with the NAS's vision model;
- gives each line the photos of that spot, with a suggestion: mark visible, no such mark visible, or cannot tell;
- shows whether each photo was already on the NAS before the landlord wrote;
- says "no photo found" only after every photo, and every photo of that room, has been checked.
The tenant decides every row. Photos that prove the landlord right are shown too. With the tenant's approval, the sheet, an album and a factual draft reply are saved on the NAS. Nothing is sent, deleted or moved.
How I built it
The Skill is Markdown only, built with UGREEN's uliya-skill-kit 0.7.7. It delegates to four registered experts: document-agent, photo-agent, file-agent and knowledge-agent. It uses catalog actions such as ug-photo search-picture, vision-vlm and search-screen. It handles every task state, asks before every write, and passes the kit's checker with 0 errors and 0 warnings.
AI tools:
- Claude Code (Claude Opus 5.5) helped read the rules and kit, and drafted and checked the Skill and this text.
Challenges
Making absence honest. A vision label can miss a door at the edge of a frame. So "no photo found" needs full coverage plus a room-by-room check, and the test set includes exactly that trap.
What's next
On the UGREEN AI NAS testing day, I'll run the synthetic test flat and report the results as measured: label accuracy, run time, whether the local model runs, and whether vision accepts photo-library paths.
Built With
- markdown
- ugos
- ugreen-ai-nas
- uliya-skill-kit