Inspiration

Our phones back up thousands of screenshots to the NAS, and the useful information is buried in them: a lab deadline in a group chat, a lecture slide you want to revise from, a booking reference, the same screen captured five times. Nobody sorts these by hand. We wanted an AI Skill on UGREEN AI NAS that does the sorting and keeps sensitive screenshots out of sight. It fits the Automate Your Workflow and Bring Memories to Life tracks.

What it does

Screenshot Rescue cleans up and sorts a screenshot folder on the NAS. It:

  • reads the on-screen text of each new screenshot;
  • sets exact and near-duplicates and junk aside in a Review folder;
  • sorts everything else into Deadlines, Lecture Slides (one folder per module code, such as CS2040), Receipts and Tickets, Info, Chats and Other;
  • flags screenshots with one-time codes, bank screens or ID numbers, and offers to move them to a Private folder;
  • on request, builds a dated deadline list and a calendar file (.ics);
  • on request, builds flashcards and a quiz from lecture-slide text, naming the source screenshot for every item.

Nothing moves until the user approves the plan. Nothing is deleted except from Review, and only after the user confirms the exact number of files. Every run writes a report that can be used to undo it.

How we built it

The Skill is a set of Markdown instructions written for the UGOS Skill Developer Package and the Uliya Skill Development Standard. SKILL.md holds the frontmatter, the goal, the required inputs and the orchestration. Five reference files cover the category rules, deadlines, the study kit, safety and confirmation, and disclosures.

The parent Skill never touches the NAS directly. Every operation is delegated to one of three registered experts:

  • file-agent (ug-filesystem, ug-filemgr, file-workflow) for finding, tagging, moving and deleting files;
  • photo-agent (ug-photo, vision OCR and VLM) for reading screenshot text and describing each screen;
  • knowledge-agent (ug-knowledge) for saving reports, deadline lists, calendar files and study materials.

Work runs one stage at a time, in batches of 50 screenshots. The parent does the reasoning over real returned results: classification, sensitivity checks, duplicate grouping, date resolution and writing flashcards. Repeat runs use an sr-processed tag, so each run only handles new screenshots. Sensitive files get an sr-sensitive tag. The package was checked with the uliya-skill-kit 0.7.3 static checker.

Challenges we ran into

  • Catalog limits. The NAS capability catalog has no image-similarity search, so near-duplicates are found by comparing extracted text and capture times. Two similar-looking screenshots with different text are kept as separate files on purpose.
  • Privacy versus usefulness. Screenshots often contain private chats and codes. We made the rules strict: sensitive text never goes into a chat, a report or an output file, and sensitive files are referred to by filename only. The Skill also reminds users that screenshot text is processed by the model selected in UGOS.
  • Dates. "Tomorrow" or "this Friday" has to resolve against the screenshot's capture date, not today's date. Unclear or conflicting dates go to a Needs checking list and are never guessed.
  • Two approval layers. The NAS can ask for its own approval before moving or deleting files, on top of our confirmations. The Skill treats a waiting_approval state as not done.

Accomplishments that we're proud of

  • A ten-stage workflow with explicit handling for completed, failed, cancelled, waiting_user and waiting_approval states. No file changes before the user chooses Apply.
  • A Study Kit where every flashcard and quiz question comes from the user's own slides and names its source, with no outside knowledge added.
  • A package that passes the Uliya static checker and includes its own third-party and AI-tool disclosures.

What we learned

Keeping the parent Skill out of direct NAS work was the most useful design choice. Each expert does one job, and the parent works only on real returned results, so it never invents a path or a date. We also learned that "do nothing unless approved" is much easier to design in from the start than to add later. For personal files, what the Skill refuses to say matters as much as what it does.

What's next for Screenshot Rescue

  • Run the Skill on a real UGREEN AI NAS iDX6011 Pro during Finalist testing. The Round 1 package has not been run on hardware yet.
  • Test OCR and classification on a real pile of screenshots, including Chinese-language text.
  • Tune the near-duplicate window and the sensitivity rules based on those results.
  • Offer translated flashcard sets as a standard option.

Built With

  • claude
  • file-workflow
  • icalendar
  • markdown
  • ocr
  • ug-filemgr
  • ug-filesystem
  • ug-knowledge
  • ug-photo
  • ugos
  • uliya
  • uliya-skill-kit
Share this project:

Updates

Submission history