Screenshot Vault Triage

1. One sentence: What does the Skill do?

It sorts a pile of NAS screenshots into topic folders and a masked text index after you approve the plan, and later finds one screenshot from a plain description.

2. Who is the target user and problem?

Anyone whose phone syncs to a NAS ends up with a screenshot folder nobody opens. The one you need later, a tracking number, a gate, a transfer receipt, sits among thousands of near-identical images with no useful filenames. The bank and order screens are exactly the ones you would not upload to a cloud OCR service.

3. What is your proposed solution?

Find the screenshots, read them with on-NAS OCR and vision, propose a move plan, move only after approval, then save an index and an undo log.

4. What is the Core User Flow?

You name a folder. The Skill shows a plan table; you approve or trim it. Files move, and the index and undo log are saved. Later, ask for one screenshot in plain words and get its path and date. "Undo" moves everything back. Run it again after new syncs and it handles only screenshots newer than the last index.

5. What substantive difference does your idea offer compared with existing tools or conventional AI solutions?

UGOS already lets you search text inside photos. This Skill does the part search cannot: it files screenshots by meaning only after you approve the plan, writes a plain-text index with long number strings masked, keeps an undo log so one sentence reverses a run, and picks up only new screenshots next time.

6. Which UGREEN AI NAS capabilities will be used, and at which step in the workflow?

Photo search, OCR, vision: find, classify, and answer later lookups. Document receipt extraction: amounts, if images are accepted. File workflow: plan and approved moves. Knowledge save-text: index and undo log.

7. What was your technical approach?

Markdown orchestration, no scripts. Nine steps delegate to photo, document, file and knowledge agents, each using only the previous step's real result. Work runs in batches of 100 screenshots with a progress report, and index files of 200 rows. Before saving, the parent Skill re-checks its own table and masks any run of six or more digits, whatever the classifier flagged. File changes need platform approval. Local NAS models only.

8. What is the current development status?

Done: SKILL.md, three reference files, static checker passed. Not done: any NAS run. Pending NAS validation: how well search and OCR separate the nine topics, whether the file agent accepts per-file destinations, and what batch size the device handles. If per-file destinations are unsupported, the Skill stops and says so.

9. What is your final validation plan? Name two core test cases and one edge case to validate if selected as a Finalist.

Test 1: sort 400 synthetic screenshots and compare with hand labels; the target is 85 percent agreement, the rest left unclassified. Test 2: approve, move, undo; every file returns to its original path. Edge case: a fake card-number screenshot never appears unmasked in the index, even if the sensitive flag missed it.

10. Are there any known limitations of your Project?

Classification quality is unproven on the device. Text-light or mixed-language screens may end up unclassified. Masking also hides long order and tracking numbers, which makes them harder to find. It needs an indexed photo library and working approval prompts.

AI disclosure: Claude (an AI assistant) proposed the idea and drafted the Skill and this text; the entrant reviewed and approved them. Third-party: official UGREEN Uliya skill kit (structure, checker). No other code.

Built With

• claude
• markdown
• ugreen