Inspiration
Every family archive has a shoebox problem. Scanned prints carry the day they were scanned, so decades of photos sit in the wrong year, and date search can't find them. Meanwhile about 70% of the photos people keep are never looked at again (Popsa, 2026 photo trends report).
The people who could fix both problems are the relatives who remember those years, and they sit at the same table every holiday. Nobody asks them, because filling in metadata is a chore. A game isn't.
Luis von Ahn's "games with a purpose" (the ESP Game, reCAPTCHA) showed that play can do real work. Which Year? applies that idea to one family's own archive, on a box that never sends those photos anywhere.
What it does
Which Year? is an Uliya Skill for UGREEN AI NAS. The owner says:
"Set up Which Year? for Ana, Ben and Grandma from Family Favourites, with the Scans folder."
- Setup. The Skill picks photos from different years in the album the owner named. It prepares three hints per photo with the on-device vision model; the hints never contain digits or text visible in the photo. It then creates a temporary album to show on the TV.
- Calibration rounds. These use photos with camera dates. Everyone guesses the year. The reveal shows the real date, the place, the camera model and Uliya's own guess. Points: 10 for the exact year, 7 for one year off, 4 for two or three years off, 1 for four or five years off.
- Uliya plays too. It guesses from what it sees in the photo and earns its vote the same way everyone else does.
- Mystery rounds. These use undated scans. There is no right answer, so the Skill combines everyone's guesses into a consensus year with a spread and a confidence level. Each guess is weighted by that player's measured accuracy.
- Write-back, only with approval. Confident estimates become tags, a "Dated by the family" album and a new file date. Original file times are written to a season ledger first, so a single undo request restores the original dates and removes the tags.
- Optional music round. A song from one round's year plays from the family's own music library, and everyone guesses its year.
- The season continues. The ledger carries scores and accuracy from game to game. Low-confidence photos wait in a replay queue for the next gathering.
Worked example
Calibration weights after a few games (weight = 1 / (1 + mean error in years)):
| Player | Mean error | Weight |
|---|---|---|
| Grandma | 0.5 years | 0.67 |
| Ana | 1.5 years | 0.40 |
| Ben | 4 years | 0.20 |
| Uliya | 6 years | 0.14 |
A mystery scan gets these guesses: Grandma 1986, Ana 1988, Uliya 1990, Ben 1993. The weighted median is 1988, the weighted spread is 2 years, and four guessers agree that closely, so confidence is high. Grandma's track record outweighs Ben's outlier, and the AI's opinion counts for exactly as much as it has earned.
How we built it
- Format. The Skill is Markdown only and follows the Uliya Skill Development Standard: a
SKILL.mdplus six linked references (game rules, hints and reveals, archive write-back, season ledger, disclosures, validation plan). - Orchestration. The parent Skill delegates to three registered experts and passes each step only the real results the experts return:
- photo-agent finds dated and undated photos, reads capture data, writes hints and Uliya's guess, and creates albums.
- file-agent finds the owner's home folder, logs original file times, appends the ledger, adds tags and sets file times.
- media-agent plays the music round.
- Speed. Vision work runs once at setup, so the rounds stay conversational on the local model.
- Maths. Scoring, accuracy weights and the weighted-median consensus are computed by the parent from the experts' returned data. The rules file includes the worked example above, so the model has a concrete reference.
- Scope and safety. The Skill reads only the album and folder the owner names. Every write goes through the platform approval flow, the ledger is always written before anything changes, and processing is on-device by default.
- Static checks. We wrote a local port of the kit's checker rules. The package passes with 0 errors and 0 warnings, and the port flags the violations we planted deliberately.
Challenges we ran into
- Keeping the answers secret. The host's chat may be on screen, so hints may contain no digits and no text visible in the photo. The answer key stays in the conversation until each reveal.
- A season without a scheduler. Skills can't run on a schedule or keep their own state. An append-only ledger in the owner's folder carries scores, accuracy and undo data between games.
- Approximate dates vs. exact timestamps. An estimate names a year, not a day. We use 12:00 on 1 July, which stays inside the right year in any time zone, and the photo is always labelled as an estimate.
- Safe write-back. Original times are logged before anything changes. If logging fails, nothing else is written. A declined approval is recorded, and the Skill doesn't ask again in that game.
- Outliers. An average would let one wild guess drag the estimate. A weighted median doesn't.
- Template vs. checker. The kit's template uses an "Expert orchestration" heading and
->arrows. The bundled checker expects "Orchestration", "Generation-time capability evidence" and→, and it rejects a README at the package root. We followed the checker.
Accomplishments that we're proud of
- A game that does real archival work without feeling like work.
- The AI is a contestant, not an oracle. Uliya's vote is weighted by its record, like everyone else's.
- Every write is reversible, and every estimate stays labelled as an estimate.
What we learned
- The calibration rounds aren't filler; they're the instrument. A consensus is only as good as what you know about the people in it, so measure the guessers before trusting them.
- With a local model, order matters: do the slow vision work once up front and keep play quick.
What's next for Which Year?
This is a Round 1 pre-test package. It hasn't run on a NAS yet. As a finalist we would validate, on an iDX6011 Pro:
- hint quality and speed
- the temporary album on a TV
- tag creation
- music search by year
- whether the Photos timeline follows the new file dates
After that, we'd add more modes ("Where was this?", "Who's this?" for unnamed faces, put two photos in order) and seasonal packs such as "the holidays through the years".
Disclosures
- Third-party content: none. The package is Markdown only, with no third-party code, libraries, fonts, images or data.
- AI tools: Claude Code (Anthropic, Claude Opus 5.5) was used for research, game design drafting, and writing the Skill files, the proposal and this page. The entrant reviewed and edited all content.
- Statistic: Popsa, 2026 photo trends report.
Built With
- claude
- file-agent
- markdown
- media-agent
- photo-agent
- ug-filemgr
- ug-filesystem
- ug-music
- ug-photo
- ugos
- ugreen-ai-nas
- uliya
- uliya-skill-kit