Inspiration

Picture the worst phone call of your life. Dad is in the ICU. The family is in the hallway. Who decides? Where is the will? Is it the current one? Is the insurance active?

The answers exist: a 2016 scan, a PDF called Scan0047, a folder nobody has opened since. Fewer than half of US adults have a will (Gallup, 2021). The half who do have another problem: the paper is there, and nobody knows if it still works.

What it does

Family Emergency Binder gives your papers a fire drill. It asks them the twenty questions a family faces in the first week, and grades each Ready, Look Again, Out of Date, Missing or Not Checked, with file and page.

It finds the policy that ended. The old will next to the new one. A deed and a home policy that spell the street differently.

Photo tools sort what they find. Vaults store what you upload. This tests whether what you stored still holds, then asks the one question only a NAS can ask, because only the NAS holds your papers and your photos. The will is dated 2016. Someone is in 412 photos, the earliest from 2019. The will never uses her name. The Skill does not say who she is. It asks: is this what you intended? It never gives legal or financial advice.

Your family gets a binder: who to call, where things are, every number masked. You get a fix-it list while there is still time. Run it after a move, a new baby or a renewal: it asks for a folder name and a yes or no on Photos, and a re-run shows only what changed.

The Skill itself calls no outside service. Whether the NAS's own experts run their models on the box is the first thing a real NAS test will show. A file with a long number in its name is held back and never read. Originals are never moved or changed. (Every person, file and number above is a made-up example.)

How we built it

The Skill is Markdown only: SKILL.md plus five reference files, written against the UGREEN Uliya Skill Kit 0.7.3 and its capability catalog 2.0.0. The parent delegates to four NAS expert agents (file, document, photo, knowledge) and uses find, extract, qa, reconcile, search-picture, figure-albums, tags and save-text. Three gates stop for the user before anything is written. Every write goes through the platform's own approvals and is read back to check it. The kit's checker reports zero errors and zero warnings.

Claude (Anthropic), through Claude Code, drafted the Skill, the proposal and the tests. We had no NAS during Round 1, so we built a local simulator. Real Claude Haiku models played the assistant and the NAS experts. The NAS commands, approvals and faults were simulated, run over fictional papers and photos. ReportLab and Pillow made the papers and photos, and Tesseract read the scans.

Challenges we ran into

  • Privacy. We had to keep numbers out of every place a model could see them, including the instructions sent to experts. Home addresses reached the assistant in early runs. Birth dates came through two routes. Both are now closed (0 of 27 runs on the latest round). Masking on arrival catches anything that slips through: none of these values was ever shown to the user or saved.
  • Re-runs. A refresh used to drop phone numbers from the family binder. A rule now carries them over, and none were lost in 4 of 4 refreshes on the fresh set.
  • Thin results. When the NAS returned only part of a long will, the law firm's phone went missing and one answer was wrongly marked Ready (0 of 3 runs right). Step 5 already reads every page, so it now supplies phones and sources too: 3 of 3 right.
  • Photo dates. In 1 of 30 fresh runs the photo expert misread a date and the report asked one extra question. A re-check of old dates is the next fix.
  • Honest grading. We froze the Skill before each fresh set of test papers, then hand-read the results.

What we learned

  • A checklist proves the paper exists. A drill shows whether it still answers.
  • Forced tables beat prose. Once the life check became a table with one row per person and paper, no person or paper was dropped, and the photo count and first month now come through in every fresh run that reaches that step, on both sets.
  • Repeat runs tell you more than one-off panels.
  • On two fresh, blind sets, 29 of 30 simulated runs passed, with no false Ready and nothing private ever shown or saved.
  • Not every fix is a win. A later round fixed the photo-date slip but did worse on a new blind set (12 of 15 against 15 of 15), so we kept the earlier version.

What's next

If selected as a finalist, we'll run the same drill on a real UGREEN AI NAS, on fictional papers with a known answer key. We'll record the demo on that test day.

Built With

  • anthropic-haiku
  • claude
  • claude-code
  • markdown
  • node.js
  • pillow
  • pypdf
  • python
  • reportlab
  • tesseract
  • ug-document
  • ug-filemgr
  • ug-filesystem
  • ug-knowledge
  • ug-photo
  • ugos
  • ugreen-ai-nas
  • uliya-skill-kit
Share this project:

Updates

Submission history