Inspiration

Before every trip, people who love photos save dozens of pose and composition ideas: screenshots from Instagram, Pinterest or TikTok. Then they arrive at the landmark, the light is changing, everyone is tired, and the ideas stay buried in the camera roll. The gaps only show up at home, when it's too late to reshoot.

On group trips there's a second problem: the person holding the camera often comes home with almost no photos of themselves.

Shot Scout turns those saved references into a plan, and turns the NAS into the friend who says "you still haven't got the jump shot on the Opera House steps."

What it does

Plan mode. Point it at a folder of saved references. The NAS vision model reads each image and writes a shot card: framing, camera angle, pose, light, and who should be in it. The cards are grouped by landmark and saved as a shot list, plus a Shot ideas album you can open on your phone on location.

Check mode. At the end of the day, ask "What am I still missing at the Opera House?" The Skill compares new photos with the open shot cards and reports:

  • Missing shots first, with the reference image and a reshoot tip
  • shots that need review (partial matches never count as captured)
  • captured shots, with the best take collected in a Selects album
  • which companion still has no photo at that place

Coverage per landmark is simply

$$\text{coverage} = \frac{\text{captured shots}}{\text{planned shots}}$$

and every check only reads photos added since the last one.

How we built it

Shot Scout is a pure Uliya Skill: one SKILL.md and four Markdown references, with no scripts. The workflow delegates every NAS action to three registered experts:

  • file-agent lists references, loads the shot list and report, and manages a shot-checked tag
  • photo-agent answers vision questions about each image, lists photos by folder and time, filters by named people, and creates albums
  • knowledge-agent saves the shot list and coverage report

Every write (file, album, tag) follows a preview and NAS approval, and original photos are read-only. The package passes the static check of uliya-skill-kit 0.7.7 with no errors or warnings.

Challenges

  • Designing within the capability catalog. Our first draft created albums from lists of file paths and read place names from photo listings. The official catalog doesn't support either, so we rebuilt those steps around picture identifiers, folder-based albums, and location as an optional filter.
  • No notifications, no network. A Skill runs only when asked and cannot look anything up online. We made the evening check the core habit and marked general-knowledge suggestions clearly as "suggested".
  • Large photo days. A day of travel easily produces hundreds of photos, so checks run in batches of 40 and resume through a tag.
  • Privacy. Every vision prompt forbids identifying people, and the people check only uses names the user gives for existing named albums.

What we learned

  • A good Skill is mostly careful orchestration: clear inputs, honest empty-result handling, and never claiming a write that didn't happen.
  • Checking the design against the real capability catalog early saves rewrites. Several "obvious" steps weren't possible as first imagined.
  • On-device AI changes what's acceptable to build. Comparing friends' faces and private

Built With

Share this project:

Updates

Submission history