Inspiration

I have been programming for a while and the one key habit we are always taught is that keep your .env files secret and always know where you API keys are. Most people really don't have this habit when it comes to their personal files like IDs, bank account information, old passwords when they sit on the NAS. When people back up, share or transfer their data, nobody has a record of where the sensitive files are.

I wanted to bring that same discipline to everyday files. A simple way to answer and find out "where is it?" before anything leaves the NAS, without the report exposing the content

What it does

sensitive-file-sweep is a Skill for UGREEN AI NAS. You pick folders, or run a full scan of your own personal folder, and choose whether to include photos or scanned images and code or config files. It flags suspected identifier-style numbers, financial documents, and plaintext credentials or keys, and reports each file's path, category and location. It never repeats the sensitive content.

You then choose which files to tag, and the Skill reads the tags back to confirm them. It saves a location-only report where you say. Every write needs your explicit yes. The Skill never moves, deletes, renames, overwrites or shares anything.

How we built it

The Skill is Markdown only: a SKILL.md and two reference files. A parent Skill delegates to four NAS experts (file, document, photo and knowledge) using only capabilities from the UGREEN capability catalog. Each step uses the real output of the previous step. I used Claude Code to draft and edit the files, and the Uliya static checker to check the package format

Challenges we ran into

The catalog has no content search or regex action, so detection depends on a language model reading each file. That means it can miss items or flag harmless files. Keeping sensitive values out of the results relies on instructions to the model. It is not a technical control, so I added discard rules and said so plainly. Image checks take one image per call, so images are opt-in and capped at 50 per run. UGREEN describes its AI as running on-device by default, but the capability catalog doesn’t say where each expert’s models run. So I made no claim that this Skill’s work stays on the device, and I made it a planned check.

Accomplishments that we're proud of

A complete package that passes the static checker, with every step tied to a catalog capability. Safeguards at every write: explicit consent, a check that the report path is free, a folder-permission warning, and read-backs for tags and the saved report. Honest limits: the Skill says "suspected", lists what it could not scan, and never claims a clean result means nothing sensitive exists.

What we learned

A language model reads context that fixed regex patterns miss, but it is less precise, so the report has to be careful about what it claims. Writing the limits down early made the design better.

What's next for sensitive-file-sweep

Testing on a real NAS is the main step. I have ten planned checks, such as the tag lookup format, how reliably redaction works, which file types the document expert reads, and where the models run. After that, I'd like to extend the file types covered and, if a scheduling capability becomes available, run the sweep regularly.

Built With

Share this project:

Updates

Submission history