Inspiration
I started this project knowing almost nothing about NAS hardware. A NAS is a private Google Drive that sits in your house. Once I looked at who actually uses one, the problem was obvious: a NAS fills up with unsorted receipts, contracts, and thousands of photos named IMG_1234. You can only search by filename or date, so you waste hours hunting, miss payment dates, and run out of space. The segment data made it sharper. In the consumer + SMB NAS market UGREEN serves ($3.51B in 2025), home users are the biggest slice at 54.3% and they are the ones filling drives with high-quality photos and videos. They don't want to be their own filing clerk, and they don't want their files sent to the cloud to get organized. That is what FileFlow is for.
What it does
FileFlow is an AI agent-native skill that lives on the UGREEN AI NAS and organizes everything in it, automatically.
It watches the folders you authorize, classifies each new document (receipt, invoice, contract, form), extracts the dates, amounts, and parties, describes photos and videos, and files everything into a clean structure. Files untouched for 90+ days are archived into lossless, checksum-verified bundles and the original is kept until verification passes.
The part people actually feel is search. You ask in your language for example "beach trip 2024" and FileFlow finds the photo whether it is live or archived, with archived hits marked for one-step restore. A weekly digest tells you what was filed, totals by vendor, what is waiting in review, and how much space archiving saved. Everything runs on the box. No document ever leaves the device.
How we built it
I started from the ground up: what a NAS is, what people use one for, and who the biggest customer segments are. That research shaped the whole design ,home users first, space and findability as first-class features, not afterthoughts.
The skill is specified in Markdown (SKILL.md, with workflow.md and testing.md) using the UGREEN Skill Developer Package's capabilities: File, Document, Knowledge, and System. Six specialist agents run in sequence for each new file ,Scanner, Classifier, Extractor, Fact Checker, Organizer, Summarizer. A pattern I had already used in an 8-agent AI job-search system I built, where small agents each do one job and a Fact Checker verifies the result before anything is committed. Search is query-time, not a seventh agent: the Classifier writes AI descriptions and the Organizer records archive locations in the knowledge index at ingest, so search matches a natural-language query against that index.
Challenges we ran into
The biggest one was building for hardware I don't have. Round 1 needs no NAS, but it meant I could not verify exact CLI command names or test anything end to end .The Developer Package docs were unreachable when I started, and the Starter Kit page only became reachable later.
The 500-word proposal limit was the second fight. Fitting the Archiver and the AI search into it meant cutting everything that wasn't essential, and resisting scope creep .An early idea for a compression agent died when the math showed documents are tiny and photos/video are already compressed, so the Archiver archives instead of compressing.
The subtlest problem was findability. A plain archive makes files invisible to search, so the design had to keep an index of every archived file's location and description from the moment it was ingested. And AI descriptions can be wrong, video search needs frame sampling, and local compute is limited .So low-confidence items go to a review folder instead of being filed on a guess.
Accomplishments that we're proud of
A complete, coherent skill package in 496 words: a six-agent pipeline with a Fact Checker gate, an Archiver that never deletes an original before checksum verification, and search that covers live and archived files together. I'm proudest of the safety defaults , nothing is filed, moved, or deleted on a guess, and unreadable or encrypted files are flagged, never deleted. The design also survived its own trade-off checks: the Archiver and the AI search were added only after weighing their costs, not because they sounded impressive.
What we learned
I learned what a NAS is for, who uses one, and why the home segment matters most for this product. I learned that the boring constraints , a 500-word limit, no hardware to test on, a local model instead of the cloud ,are what force a design to be clear. And I learned that for a tool that moves and deletes people's files, trust is the feature: verification before filing, originals kept until verified, and uncertainty routed to a human instead of hidden.
What's next for FileFlow
If FileFlow reaches the finalist stage, the plan is to test it on a UGREEN AI NAS iDX6011 Pro during the Nov 22–28 remote-access window, following the test plan in testing.md: 30 mixed documents (at least 27 classified and filed correctly), a digest that lists every file and the space saved, a search test that finds live and archived matches, and edge cases (a low-resolution scan and a password-protected PDF) that must be flagged, never misfiled or deleted.
After that: build the agent code and CLI integration, refine the filing rules and AI descriptions from real test results, and keep the promise the proposal makes ,a NAS that organizes itself, and lets you find anything by meaning.
