Library Atlas crawls one user-authorized NAS directory and produces a searchable, provenance-cited index: the atlas. It reads documents, groups photos and videos, transcribes recordings, and flags duplicates. Every fact in the atlas cites its source file. After the atlas is saved, the skill offers an optional tidy-up that moves and tags files into an organized archive. Nothing moves without explicit approval.
The problem: NAS libraries grow faster than their owners' memory of them. Thousands of files across years, no reliable index, and an assistant answering from manual uploads instead of the real library. Finding one contract, photo, or duplicate means digging by hand.
How it works: the user names and authorizes a directory and scope. The skill crawls the tree in batches. Expert agents read documents, scan duplicates, group photos and videos, and transcribe audio. The atlas is saved. The user reviews a summary and approves or declines the tidy-up. Approved moves are reported per item.
Capabilities used: Files (crawl, step 1; classify and scan duplicates, step 2; confirmation-gated tidy-up, step 8), Documents (extract metadata, step 3), Albums (photo groups, step 4), Media (video shortlist, step 4b), Knowledge (transcribe audio, step 5; save the atlas, step 6).
Technical approach: SKILL.md orchestrates registered experts through supported UGOS capabilities: file-agent crawls with ug-filesystem find, ls, stat; document-agent extracts dates, parties, titles, totals with ug-document extract; photo-agent groups with ug-photo figure-albums, search-picture, timeline-summary; media-agent shortlists videos with ug-video search and shortlist; knowledge-agent transcribes with audio-transcription and saves with save-text; approved moves use file-workflow batch-move and ug-filemgr tag-add-file. Inputs: directory, scope, output location. Permissions: read-only; writes only to the atlas and approved moves. Local model only; no external calls. The skill never reads credential-like files and extracts no secrets or personal identifiers.
Status: the complete SKILL.md workflow, reference guides, and failure handling are built, written against the Starter Kit capability scope. Nothing structural remains unbuilt. NAS testing is pending: end-to-end runs on real libraries, batch-size tuning, confirmation-flow behavior on-device.
Validation plan: Core 1: a mixed directory (50 documents, 200 photos, 5 recordings) produces a complete atlas with correct provenance per fact. Core 2: the tidy-up moves only approved items and reports each move. Edge: a directory containing unreadable and credential-like files completes: unreadable files are listed with reasons; credential-like files are skipped and counted, never read or named.
Limitations: on-demand snapshots, not continuous monitoring; the atlas goes stale as files change, re-run to refresh. Very large libraries need chunked runs. Extraction quality depends on the local model; encrypted or password-protected files cannot be read. Depends only on supported UGOS capabilities and the on-device model; no external services or third-party components.
AI disclosure: an AI assistant (Muse Spark, by Meta) drafted the orchestration text and reference documents, working from the official Starter Kit, the official capability catalog, and the SkillHub content standards. The official UGREEN creation tool (uliya-skill-kit, version 0.7.7-en) ran the static checks. The entrant reviewed the drafts and remains responsible for the submission.
Built With
- on-device-local-model
- ugos-cli-capabilities
- ugos-skill-creation-tool
- uliya-skill-development-standard
