media-finder
UGREEN Hackathon 2026 | Round 1 Project Proposal | Pre-test Skill 0.4.0
1. What does the Skill do?
media-finder helps incoming media coordinators and creative teams retrieve NAS assets and propose evidence-backed tags without knowing the archive layout.
2. Target user and problem
A new coordinator inherits campaign photos, intro videos, Illustrator designs, and recordings. Previous staff know where these belong; the replacement knows the event or intended use, but not filenames.
3. Solution
Connect observed folder context, native tags, indexed photo evidence, and explicit correction records. Return original paths with match reasons, separating confirmed matches from related candidates and unsupported contents.
4. Main user flow
Ask for “2026 summer campaign materials.” Discover accessible shared metadata, prioritize observed event/year matches, and report verified candidates after a productive batch, with remaining coverage. Continue from known pending paths when requested. Confirm tag proposals before writes and verify resulting associations.
5. Differentiation
Test against ordinary filename/tag search: can event-folder context connect an obscurely named asset, can indexed OCR find a brand logo, and can a successor reuse a confirmed correction? The contribution is evidence-linked handover across media; improvement is not yet demonstrated.
6. NAS capabilities and workflow steps
file-agent uses ug-filesystem/ug-filemgr for discovery and tags; photo-agent uses ug-photo for indexed OCR/categories and supported scene search. knowledge-agent reads correction records or transcribes explicitly requested audio. document-agent saves authorized records through ug-document.
7. Technical approach
A Markdown Skill delegates scoped tasks to registered experts. Shared metadata is the default; user bounds prevail. Folder-filtered photo searches need verified photo-service identifiers. Tag writes require platform approval and read-back. Platform experts/models supply inference; no custom model or external service is added. Actual model routing and data handling need device confirmation.
8. Current development
Version 0.4.0 defines request-driven discovery, evidence-based proposals, and correction sequencing. Current static checks pass; synthetic workflow checks cover the previous version. New prioritization/stop rules need behavioral and runtime validation. No NAS execution or measured retrieval quality is claimed. Video/design contents are generally metadata-only; supported photos additionally use indexed evidence.
9. Finalist validation plan
Core one: on 20 authorized assets with predetermined target paths, compare filename/tag search with the Skill; target at least 8/10 relevant paths retrieved and zero unrelated results labeled confirmed. Record latency. Core two: replace a tag, verify only intended associations changed, then have a second account retrieve three expected assets using an accessible correction record. Edge: deny a subtree; require no access bypass or false completeness. These are planned acceptance targets, not results.
10. Limits, risks, and dependencies
Depends on UGOS permissions, indexes, and expert availability. Preview limits can prevent complete retrieval. No arbitrary video-scene analysis, universal AI/PSD rendering, or unattended tagging. Cross-account native tags and large-archive performance remain unverified.
AI assistance and components
AI assistance: OpenAI Codex supported workflow design, drafting, evaluation tooling, and review; Aside supported official-source research. Reference: UGREEN Uliya Skill Developer Kit 0.7.7, under its supplied terms. No third-party code, models, or media are bundled in the Skill ZIP.
Official references: https://ugreenhackathon.devpost.com/ ; https://ai.ugreen.com/pages/hackathon-starter-kit
