Why this exists

A photo library can find a beach picture. It cannot ask the person who was there to name an unknown face. Before the Visit prepares that conversation while there is still someone to ask.

What it does

Before a visit to an older relative, the Skill proposes up to twelve family photos and one question for each. It starts with photos showing the relative beside an unnamed person, then follows unanswered questions and names from earlier visits. The user reviews the selection before a new album and question sheet are created.

After the visit, the user supplies a recording made with the relative's agreement. The Skill matches the transcript to the questions, shows the actual words and timestamps for review, and saves only approved answers. Unanswered questions stay open. The next visit builds on those notes.

Technical approach

The five-file Markdown package documents a workflow using photo-agent, knowledge-agent and file-agent through ug-photo, ug-knowledge and ug-filesystem. Each step uses the real output of the previous step. Searches stay within user-approved folders. Saving requires approval; empty or failed results are reported instead of filled in.

The Skill cannot assign names in Albums. It returns a list for the user to name manually. No outside model or service is specified. The actual locations of vision and transcription processing still need verification on the NAS.

Round 1 status

SKILL.md and four linked reference files are complete. The original package passed the Creation Tool's static validation. The submitted revision changes documentation and disclosure; that revision has not been re-run with the official checker. Nothing has run on a UGREEN NAS yet, and no working-hardware demonstration is claimed.

Finalist test plan

On a public-domain test library, propose twelve photos led by two-person matches and create an album only after confirmation.
File five approved answers from a recording, then leave those photos out of the next run.
If a recording answers no questions, save nothing and leave the questions open.

Face grouping, scanned-print detection, older voices, and local processing all need hardware testing.

Development and disclosure

Created during this Hackathon using the UGOS Skill Creation Tool documentation and capability catalog. Claude Code drafted the workflow, references and proposal; OpenAI Codex reviewed the materials and prepared the documentation revision and submission files. No third-party components are bundled. Full disclosure is included in the ZIP.