Why this is new

Google Photos and Apple Photos already recognise faces, places and objects, and resurface "on this day" memories. But they only know what is in the pixels: they can tell you there is a beach and two people, never who held the camera or why everyone was laughing. Photo Question collects the part only humans know, one tiny question a day, and lets the whole family answer the same photo, so the library slowly becomes a searchable archive of stories instead of a pile of files.

The data it runs on

Source: the user's own photo library. On iPhone through PhotoKit with "selected photos" or full access; on Android through the MediaStore API; or a Google Takeout export (photos plus JSON metadata with date and GPS) for people who want to bring years of cloud photos.

What we read: date taken, GPS place (reverse-geocoded with the open Nominatim or OpenStreetMap data), face clusters already on the device, and a small thumbnail for the vision model. Originals never leave the phone unless the user shares a story with family.

Limits and gaps: scanned prints have no date or GPS, so the first question for those is "roughly what year?". iOS limited-access mode only shows photos the user picked. Face naming is not available in every region (biometric rules in Illinois and the EU), so the app asks "who is this?" in words instead of auto-tagging.

Fallback: a "shoebox" mode: point the camera at old prints, one per day.

1. The mission

Every family photo has a story that lives in one person's head. Photo Question asks for that story one photo at a time, before it is forgotten, and keeps it next to the picture forever.

2. Why people can't resist it

• One question is easy. Captioning 4,000 photos is impossible. Answering one question with your morning coffee takes 40 seconds.
• Curiosity gap. A good question ("what was cooking in that pot?") pulls a memory you didn't know you still had.
• The family chorus. Seeing your brother's version of the same day is irresistible, and he can't resist correcting you.

3. How it works for real

1. Pick: a scoring pass favours photos with people, older dates and no story yet, avoiding screenshots and duplicates (on-device heuristics, no cost).
2. Look: a vision model (Gemini Flash or GPT-4o mini via OpenRouter) describes the thumbnail: setting, objects, mood.
3. Ask: a frontier LLM (Claude or GPT) writes one specific, answerable question from the description, date, place and what the family already said. It avoids anything the photo can already answer.
4. Listen: voice answers transcribed with Whisper or a Gemini audio call, in any language, kept as audio too.
5. Write: the LLM turns the answer into a short caption and a story in the user's own words, plus tags (people, place, food, feeling). Embeddings of each story (a small embedding model) power search by memory.

4. Cost per use

Pennies per day: one vision description (about $0.001), one question (about $0.002), transcription of a 40 second answer (about $0.004), caption and story (about $0.003). Under $0.01 a day per user, about $0.30 a month.

Free: one question a day, your own library, search. Family ($3 a month): up to 12 relatives answering, shared album, audio kept forever, and a printed "year of stories" book at cost.

5. Growth loop

Every story ends with "Ask Ana and Dad". Relatives answer through a link without installing; each answer shows their own photos would get questions too. Family members install to get their own daily card.

6. Risks and guardrails

• Privacy: originals stay on the phone. Thumbnails are processed and deleted. Nothing is shared unless the user taps share.
• Grief: questions about people who have died are gentle, and the user can mark a person "ask carefully" or "don't ask about".
• Sensitive photos: a filter skips medical, intimate, document and screenshot images before any question.
• Disagreement: family versions sit side by side. The app never decides whose memory is "true".

7. MVP build order

1. Photo picker (PWA file input, then native) with EXIF date and place.
2. Vision description plus Claude question prompt, one card a day.
3. Voice or text answer, caption and story generation.
4. Library with "has a story" ribbons, streak calendar and keyword plus embedding search.
5. Family answer link (no install) and the side-by-side versions.