Narrow Photo Search

A UGOS Skill that turns a too-long photo search into a short conversation: one question at a time, with counts from the real results, until a few candidates are left.

Comment

Inspiration

Two things I said when this started, translated from Chinese. First: "Categories make it hard to find the one I want most." Then, about searching the photo album on my phone: "It just gives too many, and I have to swipe through them one by one. It isn't precise. It feels like it needs more hints to be more precise."

What it does

You describe the picture you remember. The Skill searches the UGOS Photos library with the details the library can search. If too many items remain, it asks one question whose answer splits them best, such as year, month, place, or person, and shows the real count next to each option, plus "not sure". Details no index field can hold, such as what people were wearing, are checked by the vision capability (vision-vlm), but only when 30 or fewer candidates are left; items it rejects are counted and can be shown again. You get up to six candidates, each with the reason it matches, and the number of items left after each step, for example 312 → 88 → 28 → 5. Saving them as an album is optional and goes through approval.

How it is built

  • UGOS Skill Creation Tool 0.7.3 and capability catalog 2.0.0.
  • One expert, photo-agent (ug-photo). The parent Skill only delegates, passes on only completed results, and handles all five delegation states.
  • Read-only by default. timeline was left out because it changes data.
  • Where vision-vlm runs is not documented in the catalog, so the Skill asks photo-agent for a model on the NAS and asks the user before any picture could go to a cloud model.
  • Official static checker: 0 errors, 0 warnings. Nothing has run on a NAS yet.

Challenges

The official template's section heading ("Capability evidence") and the arrow format in the kit's own instructions ("->") both fail the kit's static checker, which expects "Generation-time capability evidence" and "→" between backticks. Reading the checker and testing both variants showed it before submission.

What's next

If selected: test on the iDX6011 Pro with a seeded library of licensed photos and videos (two core cases and one edge case, as in the PDF), then extend to documents and recordings.

AI use

Claude Code with Claude Opus 5.5 (Anthropic), and Claude Sonnet 5.5 for reviews. Mine: the problem, the quoted words, and the idea of narrowing with more hints. Claude Code: design details, limits, working name, Skills Hub search, Skill files, checker runs, reviews, translation, the PDF answers, the text on this page, layout. Third-party components: none. External services: none called by the Skill; if vision-vlm offers only a cloud model, the user is asked first.

Built With

Share this project:

Updates

Submission history