Inspiration
Most people find out how weak their insurance claim is at the worst possible moment: right after a fire, a break-in or a flood. The receipt is gone, nobody photographed the serial number, and the policy turns out to exclude the one item that mattered. The payout gets cut or denied.
Looking at this hackathon's Automate Your Workflow entries, I saw the same pattern over and over: tools that assemble a claim file after the loss. I wanted to attack the problem from the other direction. What if your NAS could tell you, today, which of your belongings would fail a claim, and exactly what to fix?
A NAS is the right place for this. Receipts, home photos and insurance policies are some of the most private data a household or small business owns. Auditing them is only credible when none of it leaves the device.
What it does
Claim Readiness Auditor scans the folders you choose and builds a register of your insured assets from real receipts, warranty documents, policy documents and photos. Each asset gets four checks worth 25 points each:
| Pillar | Check |
|---|---|
| Proof of ownership | Receipt matched to the asset |
| Proof of condition | Dated photo of the asset |
| Proof of condition | Legible serial number |
| Coverage | Policy clause that covers it |
$$ S = c_{\text{receipt}} + c_{\text{photo}} + c_{\text{serial}} + c_{\text{coverage}}, \qquad c_i \in [0, 25] $$
Partial evidence earns partial credit. For example, a photo without a capture date earns 10 points instead of 25. Assets land in a band: High ($S \ge 80$), Medium ($40 \le S < 80$) or Critical ($S < 40$).
The output is an Evidence Gap Action List, ranked by impact:
$$ P = \Delta S \times w(v), \qquad w(v) = \begin{cases} 3 & v \ge 2000 \ 2 & 500 \le v < 2000 \text{ or unknown} \ 1 & v < 500 \end{cases} $$
where $\Delta S$ is the number of points an action would recover and $v$ is the asset's value. So a missing serial photo on an expensive espresso machine ranks above a missing receipt for a cheap lamp.
In the demo, a café espresso machine starts at 25%: covered by the policy, but with no receipt and no serial photo. One dated photo of its serial plate earns both photo checks, and a re-run shows 75%.
How I built it
I built the Skill with the official Uliya Skill Kit (0.7.3). The Skill is a single SKILL.md plus two linked references: a scoring rubric and a report format. It never calls NAS capabilities directly. It delegates each step to the registered expert that owns the capability:
file-agent(ug-filesystem) finds receipts, warranties, policies and earlier reports.document-agent(ug-document) extracts items, prices, serial numbers and policy terms.photo-agent(ug-photo) finds photos of each asset, reads EXIF dates, and runs OCR on serial plates and receipt images.- The Skill merges everything into an asset register (same serial number, or same brand and model, means the same asset).
document-agentchecks coverage for each asset and quotes the supporting clause.- The Skill scores each asset and ranks the fixes.
- Optionally,
knowledge-agent(ug-knowledge) saves the report. This is the only write, and it needs the user's approval.
Every step except the optional save is read-only. Every finding traces back to a real file. The package passes the kit's static checker with 0 errors and 0 warnings.
Challenges
- Thinking in delegation, not code. My first instinct was to "call
ug-photo". The Uliya model doesn't allow that: the parent Skill can only hand self-contained instructions to experts and pass their real results downstream. I had to rewrite every step as a brief that an expert could carry out with no other context. - A strict standard. The checker enforces a lot: only
SKILL.mdandreferences/in the package, exactly two frontmatter fields, a specific capability-mapping format, single-quoted instructions, and explicit handling of every delegation state (completed,failed,cancelled,waiting_user,waiting_approval). Getting to a clean pass took several rounds. - Never inventing data. A scoring tool is tempting to "fill in". I designed every step so that an empty result is reported as empty, a failed step marks its checks as not checked rather than missing, and uncertain asset matches are flagged instead of merged.
- Making the score explainable. Weighted formulas looked clever but were hard to explain. Four equal 25-point checks let a user see at a glance why an asset scored 25% and what will move it to 75%.
What I learned
- Strong agent workflows come from scoping each step tightly, not from writing longer prompts.
- Honest failure handling (partial results, explicit states) is a feature, not boilerplate, especially when the output is a judgment about someone's finances.
- Local AI is more than a privacy checkbox. For an audit of receipts, photos and policies, it is what makes the product trustworthy at all.
What's next
If selected as a finalist, I'll validate the Skill on a real UGREEN AI NAS. I'll test receipt extraction, serial-number OCR and coverage answers on a deliberately messy household folder, then tune the rubric based on what the experts actually return.