Dossier Readiness

Inspiration

When I moved to France, I discovered that renting a home starts with paperwork. Every landlord or agency asks for a "dossier": an ID for each tenant, the last three payslips, an employment contract or a recent employer letter, the latest tax notice, a proof of address less than three months old. Buying a home means the same exercise with an even longer list.

The documents always existed somewhere: in a NAS folder, in Downloads, in a photo taken with my phone. But they kept getting lost, and every new application meant another evening of searching, checking dates and renaming files. In a market where the first complete dossier wins, that time really matters.

What it does

Dossier Readiness turns a checklist into a ready-to-send dossier, built only from documents that are already on the NAS.

1. The user names the dossier type (rental application, mortgage application, or their own checklist stored on the NAS), the folders to search and the applicants.
2. The skill finds candidate documents, including photographed paper documents found through text recognition. It reads each one to decide what it really is: a file called scan_0034.pdf can be a payslip.
3. It checks every item. Is the ID still valid? Is the proof of address recent enough? Are all of the last three months of payslips there? Do names and addresses match across documents?
4. It shows a gap report with a status for every item (Ready, Missing, Outdated, Expired, Inconsistent or Needs review), the evidence behind it and the next action.
5. Only after the user confirms, it copies the selected documents into a new folder, renamed in checklist order, and saves an index. Original files are never moved or changed.

How I built it

A UGREEN Skill is pure Markdown: a SKILL.md that orchestrates registered NAS experts. I started from the official capability catalog in the Skill Developer Package and mapped every step to a real action:

• file-agent lists and copies files;
• photo-agent reads text from phone photos;
• document-agent extracts fields and reconciles names and addresses across documents;
• knowledge-agent saves the index.

Requirement lists and validation rules live in separate reference files, so a new dossier type is just a new Markdown file. Every delegation states what happens on empty, partial, failed, cancelled and waiting-for-approval results.

To prepare validation, I generated a synthetic SPECIMEN dataset of 17 documents for a fictional couple, with deliberate traps: an expired ID card photographed on a table, a missing month of payslips, a name variant, a duplicate file, a misnamed scan, an outdated bill and a stranger's payslip. I used Claude as a pair to draft and review the skill files and the test data. The official checker passes with no errors and no warnings.

Challenges I ran into

• No NAS in Round 1. I designed against the capability catalog only and kept a clear list of what must be validated on the real device.
• Markdown only, no scripts. All the logic has to be precise instructions to experts, and each step can only use the real results of the previous one.
• Trust. Dates must come from the document content, never from file dates, and a missing document must never be filled with something that only looks similar.
• No ready-made extraction schema for payslips or ID documents. The instructions name every field explicitly instead.
• Language. The on-device AI is optimized for English, German, Japanese and Chinese, so the first version and the demo are in English.

Accomplishments that I'm proud of

• The skill only reads until the user confirms, and then it only copies.
• The gap report explains every status with evidence, not just yes or no.
• It handles two applicants, photographed documents and the user's own checklist.
• The test dataset covers the mistakes that are easy to make under time pressure: an expired ID, a forgotten month, a file with a meaningless name.

What I learned

The hardest part of a document workflow is not finding files but knowing when to stop: "this is missing" is more useful than a confident guess. I also learned how much a clear gap report is worth. It turns a stressful evening into a short to-do list.

What's next for Dossier Readiness

• Validate the full flow on a real UGREEN AI NAS and tune the instructions with real results.
• A French "dossier de location" reference based on the official list of documents a landlord may request, then other countries.
• Guarantor support and more dossier types: school enrolment, residence permit renewal, job applications.
• French and other languages as on-device model support grows.
• Warnings before a stored document expires, once the platform supports scheduled checks.

Built With

• ai
• claude
• nas
• skill
• ugreen