German Letter Decoder: UGREEN AI NAS Skill proposal
1. One sentence: What does the Skill do? It turns scanned German letters from authorities and companies into plain-language cards in the reader's own language and a deadline-sorted action board, entirely on the NAS.
2. Who is the target user and problem? People living in Germany who do not read bureaucratic German comfortably: international students, immigrants, expats, and natives who dislike Amtsdeutsch. Letters from the Ausländerbehörde, Finanzamt, Krankenkasse or Jobcenter carry deadlines; missing one means fines, lost benefits or residence trouble. Today people paste private letters into public chatbots or ask friends.
3. What is your proposed solution? A Skill that reads the letters a user already scans into a NAS folder and answers, per letter: who sent it, what it wants, the exact deadline wording, amount and payment reference, consequences, and requested documents. A follow-up question about one letter is answered only from that letter, with the location quoted.
4. What is the Core User Flow? "Decode the letters in /Scans/Post this week" → files listed → fields extracted → deadline sentences located and cross-checked → cards in the user's language → action board → optional Markdown digest → optional filing by sender and year after the user approves a full move plan.
5. What substantive difference does your idea offer? Generic translators translate words; this Skill finds what to do and by when, refuses to guess dates, reports relative deadlines verbatim, and flags conflicts. Sensitive letters never leave the device, unlike cloud chatbots.
6. Which UGREEN AI NAS capabilities are used, and where? ug-filesystem find/stat (listing), ug-document extract and qa (fields, deadline sentences, follow-ups), ug-document translate/writing (cards), ug-photo vision-ocr (fallback for unreadable photos), ug-knowledge save-text (digest), ug-filesystem mkdir and cpmv (approved filing).
7. What was your technical approach? The parent Skill only delegates to file-agent, document-agent, photo-agent and knowledge-agent with self-contained instructions; each step passes only verified results to the next. Inputs: folder, language, optional period, digest and filing folders. Limits: 10 letters per run, no overwriting, every write and move needs approval. All states (completed, failed, cancelled, waiting_user, waiting_approval) are handled. Data stays on the NAS.
8. What is the current development status? Implemented: complete SKILL.md, a 34-term field guide, README; passes the official uliya-skill-kit 0.7.3 static checker with 0 errors and 0 warnings. Pending: validation on a real iDX6011 Pro.
9. What is your final validation plan? Test 1: five mixed letters (PDF scan, HEIC photo) produce correct senders, deadlines and an ordered board. Test 2: approved filing moves files into sender/year folders without overwriting. Edge case: a letter with only "innerhalb eines Monats nach Bekanntgabe" must produce no invented date.
10. Are there any known limitations? Not legal advice. OCR quality depends on the scan. Relative deadlines need the user's delivery date. Ten letters per run. Translation quality follows the on-device model.
AI use: Claude (Anthropic) assisted with drafting the Skill text and this proposal; all content was reviewed by the author. No third-party services or data flows beyond the NAS.
Built With
- markdown
- ugreen-ai-nas
- uliya-skill-kit