Inspiration

We screenshot coupons, event tickets, booking confirmations, receipts, and addresses "for later", and then forget them. Coupons expire and bookings are missed because nobody reviews the screenshots that pile up in the phone backup.

What it does

Screenshot Keeper is a Skill for UGREEN AI NAS with two modes.

  • Review mode: for a time range the user confirms, it collects screenshots from the NAS photo library, reads and classifies each one (Coupons, Tickets and bookings, Receipts, Reference notes, Other), extracts every date with its label ("valid until", "check-in"), and shows a digest: due in the next 7 days, later, past, no usable date, sensitive, and unreadable. Screenshots that show card or account numbers are flagged without repeating the numbers. After approval it saves the digest as a text file and creates dated albums by type.
  • Find mode: locates a screenshot by the text on it, such as "the restaurant address".

It never deletes, moves, or edits the original images, and it requests no external service.

How we built it

The Skill is a SKILL.md orchestration document built with the official UGOS Skill Creation Tool (uliya-skill-kit 0.7.3). It delegates to photo-agent (ug-photo: search-screen, meta, vision-ocr, vision-vlm, albums, create-album) and knowledge-agent (ug-knowledge: save-text). Each delegation is self-contained, because child experts do not see the conversation, and each step uses only the previous step's real result. It passes the kit's static checker with no errors or warnings.

Challenges

  • search-screen has no time-range parameter, so the Skill reads each screenshot's capture time with meta and falls back to upload time, marked as such.
  • There is no add-to-existing-album action, so albums are created new with the review date in the name.
  • Keeping sensitive numbers out of titles and the digest, and handling ambiguous dates such as "03/04" without guessing.

What we learned

Writing for an agent that delegates to other agents is mostly about making every hand-off explicit: values, empty results, partial results, and every delegation state.

What's next

Testing on a real NAS: screenshot recall, the capture-time filter, OCR on dense screenshots, and date extraction.

Details are in the attached proposal PDF and Skill package ZIP. AI tools used: Claude (Anthropic) via Claude Code, to draft and check SKILL.md and the proposal.

Built With

Share this project:

Updates

Submission history