Couples Cut

Inspiration

After our Indian Hindu wedding, my wife and I had thousands of photos and one question: which ones go in the album? We had to dig through every folder to find the mehendi, the pheras, and every other ritual, curate a folder by hand, and we still worried we had missed someone.

Existing tools did not help. Auto-albums collect every match, "best photo" pickers optimize for looks, and none of them know what a phera is or that a wedding can span several days. What we really needed was simple to say: make sure every ceremony and everyone we love is in the album.

What it does

Couple's Cut is a Uliya Skill for UGREEN AI NAS that turns a wedding photo folder into an album that provably covers every ceremony and every important person.

1. Segment: you say when each ceremony happened, in your own words ("mehendi Friday evening, pheras Saturday 15:00–18:00"). Photos are grouped by real capture time. It works for any tradition (pheras, a nikah, a chuppah, a first dance) because nothing is hard-coded to one religion.
2. Measure: a coverage report shows what is missing, for example "Pheras: 4 of 120 available, Vidaai: 0, Grandma: 1."
3. Steer: you set a Coverage Contract in one sentence ("at least 15 from the pheras, Grandma 3 times, lock 5 and 12"), and targeted searches fill the gaps.
4. Save: after you approve, it creates a native UGREEN album, with an optional copy of the originals into a new folder.

Formally, for a shortlist $S$, the photos $P_c$ of ceremony $c$, and the photos $P_p$ of person $p$, the contract is a set of minimum counts:

$$ |S \cap P_c| \ge k_c \quad \text{for every ceremony } c, \qquad |S \cap P_p| \ge k_p \quad \text{for every person } p $$

If a goal cannot be met from the real photos, the Skill says exactly how close it got instead of pretending.

How I built it

• UGOS Skill Creation Tool 0.7.3 and its capability catalog 2.0.0 from the Hackathon Starter Kit. The Skill is a Markdown-only orchestration Skill: a SKILL.md plus two linked reference files, with no scripts or binaries.
• Registered experts only: photo-agent for ug-photo (search-picture by folder, time and person, figure-albums, search-screen scene labels, create-album) and file-agent for ug-filemgr scan-dedupe and the optional ug-filesystem export. The Skill never calls a CLI directly.
• Self-contained delegations: experts cannot see the conversation, so every instruction carries the exact folder, ceremony times and photo identities returned by the previous step.
• Safety by design: album creation and file copies wait for approval, existing albums are never merged, sources are never moved or overwritten, and all five delegation states (completed, failed, cancelled, waiting_user, waiting_approval) are handled.
• Privacy: photos stay on the NAS and no external service is used. Image analysis with vision-vlm is optional and only runs with the user's consent.
• Validation: the official checker reports a static pass with zero diagnostics.
• AI tools: I used GitHub Copilot and Claude Code to research the catalog, draft and revise the Skill and proposal, and iterate on the checker.

Challenges I faced

• No NAS in Round 1. I had to design for real result formats, paging and approvals without being able to run anything, so the Skill stops and reports honestly whenever a result is missing or ambiguous.
• Finding the right capabilities. My first version copied files into a folder. Reading the full catalog showed that ug-photo could create native albums and search by person and time, which led to a much better design.
• Recognizing ceremonies reliably. Image models may not know local rituals, and their data handling is not specified. Using capture time plus the user's own ceremony list works for any tradition, with image checks kept optional.
• Not assuming culture. Traditions vary between families, so the Skill never infers ceremonies from a religion. It only uses what the couple tells it.
• Fitting the story into 500 words for ten proposal answers.

What I learned

• A good Skill is mostly about orchestration and failure handling, not clever prompts: who does each step, what real data flows between steps, and what happens when something is empty, partial or waiting for approval.
• Measurable beats magical. "Every ceremony at least 10 photos" is something a couple can check and trust. "AI picked your best photos" is not.
• Personal photos belong on personal hardware. The NAS's local index of times, faces and scenes is what makes this possible without uploading anything.

What's next

If selected as a finalist, I will test Couple's Cut on a real UGREEN AI NAS with Hindu and Christian wedding folders: a 100-photo album where every ceremony has at least 10 photos and everyone appears at least 3 times. I will also test edge cases such as overlapping ceremonies, an existing album name and duplicate photos.

Built With

• claude-code
• file-agent
• github-copilot
• markdown
• photo-agent
• ug-filemgr
• ug-filesystem
• ug-photo
• ugos
• ugos-skill-creation-tool
• ugreen
• ugreen-ai-nas
• uliya
• vision-vlm