Inspiration

I work on flood forecasting. A forecast tells a family the water is coming, but it doesn't help them on the day after. That's when they need a policy number, proof of what they owned, and the passport they last saw in a drawer that is now underwater. Insurers tell people to keep a home inventory. Almost nobody does, and memory after a disaster is poor.

Most families already have what they need. It is scattered across scanned PDFs, phone photos of receipts and years of pictures taken in their own living room. On a UGREEN NAS all of it sits in one private place, and these are exactly the files people would never upload to a cloud chatbot.

What it does

Household Emergency Dossier has two modes.

Build mode ("get our papers ready for flood season"):

finds identity, insurance, property, medical and finance documents in the folders you allow, and extracts the fields that matter: holder, insurer, policy number (last four characters only), coverage limits, deductible, claims phone, notice deadline, expiry date;
builds a contents inventory from photos you already took, with room, item, and brand or model when a label is readable;
links inventory items to purchase receipts, so each row is labelled seen + receipt, seen only or receipt only;
writes a coverage-gap report: an expired passport, a home policy that excludes flood with no flood policy found, or receipts that add up to more than the contents limit;
shows you one plan, and after you approve it creates a dated Emergency Dossier folder with an emergency card, a document index, the inventory, the gap report and sorted copies. Originals are never moved or deleted.

Claim mode ("the basement flooded last night, what do we need?") reads the newest dossier, quotes the policy terms that apply, works out the actual deadline date, and lists the inventory items in the affected rooms with their photos and receipts. After you approve, it saves a claim checklist.

How I built it

The Skill is pure orchestration, written with the UGOS Skill Creation Tool (uliya-skill-kit 0.7.3) and Claude Code. SKILL.md hands 15 self-contained tasks to four NAS experts:

file-agent: discovery, folders, copies, tags
photo-agent: search-screen, search-picture, vision-vlm, vision-ocr
document-agent: extract, qa, reconcile, writing, translate
knowledge-agent: save-text

Each step receives only the real results of the step before it. The first phase is read-only. Writes happen only after a single plan approval plus the NAS's own approval prompts. Every one of the five delegation states (completed, failed, cancelled, waiting_user, waiting_approval) has explicit handling, including partial results. The package passes the kit's static checker with zero diagnostics.

Challenges

No NAS in Round 1. I can't yet measure how well the vision model lists items in a messy living room. So the design assumes it will make mistakes: every inventory row stays "AI-observed" until the user confirms it, keeps its source photo path, and brands come only from readable text or a matched receipt.
Useful without being dangerous. A dossier of IDs and policies is itself sensitive. Numbers are masked to their last four characters in everything the Skill writes, birth dates are never extracted, and nothing leaves the NAS.
Not inventing facts. It is tempting to fill gaps with plausible values. The Skill records "not found" instead, and claim answers must quote the policy or say "not stated in the policy".

What I learned

The static checker is stricter than the starter guide. It allows only SKILL.md and references/ at the package root, and it requires exact section headings. Reading the checker early saved a failed submission. I also learned that the hard part of an orchestration Skill is not calling capabilities. It is writing instructions that keep working when an upstream step returns nothing.

What's next for Household Emergency Dossier

If we reach the finals: test on a seeded share with real scans and room photos, measure inventory accuracy, tune batch sizes and the vision prompt, and record the demo.