What it does

First Seen answers “when did this first happen?” from family photos on a UGREEN AI NAS, with the earliest photo as evidence, and keeps confirmed answers in a ledger.

Who it is for

New parents like me. This summer at the beach, my wife asked whether I remembered our daughter’s first time there. We wanted to find the photos and couldn’t. I only found them a few nights ago, scrolling slowly during a break from coding. Her first year is over a thousand photos. Sitting up, standing: the same search.

The solution

I ask in my own words. First Seen searches only the album or folder I choose and shows the earliest photo beside the last one without it. I confirm and add a note. Unfound firsts stay open and resume later.

How it works

Setup message, then “go”. It reads my earlier ledger, searches what Photos recognized or checks photos oldest first until the first match, and shows an evidence card with same-day videos. I confirm, say “save” and approve; a new ledger file is written and read back.

What is different

The UGOS Baby Album sorts photos by age and suggests photos for a named milestone; its help page mentions neither the earliest one nor a date. A cloud chatbot needs my library uploaded. First Seen answers with two dated photos as evidence, not proof of the real day; resumes unfinished searches; and finds the baby by folder, never by face.

UGREEN AI NAS capabilities used

Earlier ledger: file-agent (ug-filesystem, file-workflow). Find the album, search recognized people, pets, places and objects, list photos by capture time: photo-agent (ug-photo search-picture, meta). Look at photos: ug-photo vision-vlm, local model. Save ledger: knowledge-agent (ug-knowledge save-text). Optional album: ug-photo create-album.

Technical approach

Markdown only: nine delegation templates across three experts, passing only real results downstream. Vision work is bounded: 40 photos per first per run, stopping at the first match; unfinished searches resume. The only writes, a ledger file and an optional album, need user approval; nothing is moved, renamed, deleted or overwritten.

Status

Done: SKILL.md, four reference files. The official checker passes with no warnings. Simulated: a small model given only the Skill files completed six test cases on a scripted NAS, and its commands passed the checker. Not done: nothing has run on a NAS. Pending: search sorting, recognition time, vision accuracy.

Validation plan

Synthetic set: 13 photos and 2 videos of a soft toy. Core 1, “first sat up”: stops after eight photos, answers 21 March. Core 2, new conversation: reads the ledger, finds the beach via Photos recognition, saves a second ledger beside the first. Edge: no match and a declined approval: reports not found and nothing saved.

Known limitations

The earliest photo is not the real first day. The vision model can be wrong, so I confirm every entry. Undated photos and video content are skipped. Big albums need a rough start date, or many runs. Built on catalog 0.7.3 (0.7.6 unavailable). For adults; no health judgement. External dependencies: none.

Disclosures

Third-party components: none. The package is Markdown only. External services: none; if the owner has chosen a cloud model for Uliya, the text returned by the experts passes through that model. AI tools used: Claude (Anthropic), through Claude Code, for research, drafting, the static check, a desk run, a simulation and the cover illustration; reviewed by the entrant. Test data will be synthetic.

Built With

Share this project:

Updates

Submission history