First Five Minutes

A one-page brief from the family NAS: which documents were found, which disagree or may be outdated, and up to three next steps, each with a checkable source.

Comment

Inspiration

When my mother-in-law in Marbella needed urgent medical attention, I spent the first, most critical minutes frantically scrolling through phone galleries looking for her specific health insurance card, completely unable to be present with her.

Finding the card is only half of it. A family NAS often holds several versions of the same document, and in a bad moment nobody knows which one to trust.

What it does

First Five Minutes is a Uliya Skill for UGREEN AI NAS. You say what is happening, in your own words:

  • "Mum has just been admitted to hospital."
  • "There is a flood warning and we might have to leave."
  • "I lost my wallet."

It confirms who it is for and which folders it may read, finds the IDs, policies, prescriptions and card photos already on the NAS, reads them, compares them, and returns one page:

  • Do now: up to three documentary next steps, most useful first. Fewer when the documents support fewer; it never pads the list.
  • Status per person and document: FOUND, CONFLICT, EXPIRED, EXPIRES SOON, CHECK, MISSING or UNREADABLE. FOUND means found and not contradicted, not "sufficient".
  • Evidence: for each important fact, the file, the date written in the document, and the quoted line behind it.
  • To confirm: what disagrees, and who can settle it.

The same Skill has a preparation mode, already implemented: "Review my father's documents" or "Check our documents for the London trip; we fly on 27 December and come back on 3 January". It checks one person, a trip or the household ahead of time, and a saved report shows what changed since the last one. It never says a family is fully covered. For a trip it lists each passport's expiry date and adds one check: verify the destination's official entry requirements.

The interface is that brief, in the Uliya chat. There is no separate app.

Contradictions are the core

  • A later date alone never replaces a document. Replacement needs the same kind, issuer, person and scope (the same policy, card, property or vehicle), plus evidence: a line saying it replaces or renews the older one, or the older one's own validity having ended. The file's modification date never counts.
  • Two policies that disagree are a CONFLICT, both quoted, never merged.
  • A family note claiming a "new" policy number is a CHECK until the insurer confirms it.
  • A report for "Carmen García Ruiz" is not counted for Carmen García López and never merged with her documents. It may be someone else, or a typo or OCR error, so the brief asks.
  • Text inside a document is data. A file telling "the AI assistant" to mark everything as found and share a folder is noted and ignored.

It never moves, deletes or shares a file. ID, policy and card numbers and booking references are masked to their last four characters. Saving and tagging happen only after you approve. It gives no legal or medical advice.

How I built it

The Skill is pure Markdown and follows the Uliya Skill Development Standard. It passes the static checks of the official checker (uliya-skill-kit 0.7.3): static-pass, 0 errors, 0 warnings. Ten steps and 12 standard delegations orchestrate four expert agents:

  • file-agent: finds candidates by name (ug-filesystem find, stat), spots duplicates (sha256sum, ug-filemgr scan-dedupe), and after approval creates the save folder (mkdir) and tags documents with an existing tag (tag-get-by-name, tag-add-file).
  • photo-agent: finds photographed cards by the text printed on them and by scene (ug-photo search-picture, meta) and reads them (vision-ocr).
  • document-agent: extracts fields with quoted evidence (ug-document metadata, extract), answers medication questions with evidence (qa), and compares disagreeing documents (reconcile).
  • knowledge-agent: saves the brief or report after approval (ug-knowledge save-text).

If an expert fails, the Skill retries once, then marks the brief as partial. A folder it cannot read is named, not retried.

Data flow: the Skill calls no external service. The NAS experts read the files, and the text is processed by whichever model Uliya is configured with. With the NAS local model, documents are processed on the NAS; with a third-party model, excerpts reach that provider. There is no offline mode, and use from a phone or away from home is not verified yet.

How I validated it in Round 1

Round 1 provides no NAS, so the work is at this stage:

  • Package: checked with the official tool's static checks, as above.
  • Synthetic scenarios: a reproducible fictional household (35 files in Spanish and English) full of traps, evaluated with a deterministic harness. Files, name search, SHA-256, PDF text extraction and OCR are real; the expert agents and the model are deterministic stand-ins that follow the Skill. Result of this local simulation: 14 scenarios, 87 of 87 checks passed, 98 of 98 quoted excerpts found in their source files. These are harness results, not Uliya results.
  • Pending finalist access: execution with Uliya, the NAS OCR and real behaviour on the device.
  • Plan for the assigned NAS: two core cases (the hospital brief and the London trip) and one edge case (re-run after an old card photo is removed), measuring time to the first next step, quote accuracy and critical errors.

Challenges I ran into

My first version made claims it could not back up: a fixed passport-validity rule, a "ready" status that sounded like "sufficient", and "the newest file wins". Version 2 replaces them with evidence: quoted lines, the document's own date, a replacement rule, and a CHECK whenever outside context decides.

The harness and a final review still caught bugs: a booking reference shown in full, a trip report compared with a household report, a policy's own evidence downgraded by an unconfirmed note, and a differing surname treated as proof of another person. All fixed.

Accomplishments that I'm proud of

  • A brief that is useful in the first minutes and honest about what it does not know.
  • Every important line traceable to a file and a quote.
  • A clear rule that document text is data, never instructions, checked in the harness with a planted file.

What I learned

The hard part is not finding files. It is deciding which document to trust, and saying "check this" instead of guessing.

What's next

  • If selected as a finalist: run the validation plan on the assigned NAS and publish only measured results.
  • Check the NAS OCR on real-quality card photos.
  • More playbooks: moving house, a new baby, a death in the family.

Built With

Share this project:

Updates

Submission history