Inspiration
A NAS turns into the dumping ground for every old laptop, phone and camera backup. A few years later nobody knows what is on it, which files are copies of each other, or where everything about one subject lives. Built-in tools sort by type, date or size, and search finds a filename. Nothing tells you what you actually have.
What it does
Drive Atlas maps a NAS by meaning. Every file is assigned to a plain-language project (for example "2019 kitchen remodel" or "Mom's 70th birthday") or marked unclassified. Each project gets a one-line description, dates, size, key files with real paths, and typed connections: exact copy, possible version, explicit reference, suspected same project. Every claim cites the files it rests on and carries a confidence level. The owner can then ask "what is this folder?" or "find everything about the kitchen renovation, including older versions", and the answer comes from the atlas, with paths.
It never deletes, moves or renames anything, and it never calls a copy disposable, because a copy may be the only backup.
How it works
- A parent Skill that only delegates to UGREEN's expert agents: file-agent (bounded listing, checksums of size-matched files, media info, atlas writes), document-agent (what documents are about and what they mention), photo-agent (dates, places, visible text, neutral descriptions) and knowledge-agent (short audio transcripts, evidence for questions).
- Work is done in small batches (one folder, or at most 250 files when a large folder is split by date or size). Each batch is saved as a manifest plus a progress line, so a large library is mapped over several runs and an interrupted run resumes without duplicates.
- Read-only except the atlas folder the user chooses; every write goes through the normal approval flow. The model is the NAS's local AI via its expert agents; nothing leaves the NAS.
How it was built and tested
The Skill follows the Uliya Skill Development Standard (SKILL.md plus five Markdown references, no scripts) and passes the kit's static checker with no errors or warnings.
To test it before any NAS access, a local simulator runs SKILL.md against a synthetic library (2,072 files in overlapping backups, 20 true projects, planted copies, an oversized 2,005-file folder, an empty folder, misdated files), allowing only the actions in the kit's published capability catalog, with a language model playing the parent Skill and the expert agents. A scorer compares the resulting atlas with the ground truth.
Challenges
The simulator found workflow defects that a careful read-through had missed: an oversized folder that was never recorded, a save check that accepted dispositions pointing at entries that did not exist, a resume that overwrote a valid saved snapshot, projects duplicated across batches, an undocumented date-filter code, and two folder paths that produced the same manifest name. All are fixed and retested: in the latest runs every mapped file was listed exactly once, every reported copy was byte-identical, the oversized folder was split into exact slices, and the library was never modified.
What's next
If selected as a finalist: run the same test plan on the real NAS. Three overlapping backups holding about 20 projects should give about 20 cited projects with copies counted once; "everything about the kitchen renovation" should return files from all three locations with versions labelled; and cancelling during batch 3 should resume there without duplicate entries.
Built With
- markdown
- ugos
- ugreen-ai-nas
- uliya-skill
