Incident Claim Pack

Inspiration

When something goes wrong at home, such as a burst pipe, a stolen bike or a TV that dies just after the warranty runs out, the hard part is not calling the insurer. The hard part is everything they ask for next: when exactly did it happen, photos of the damage, the invoice from two years ago, the serial number, the warranty card, and the one clause in a 40-page policy that actually matters. All of this usually already sits on the family NAS, spread across camera recordings, the photo library and a folder of scanned paperwork. People search through it under stress, while a reporting deadline is running.

A cloud assistant cannot help much here. It cannot see your cameras, your photo library and your paperwork together, and most people would not upload their insurance policy and invoices to one anyway. A NAS with local AI can do all of this privately, and that is why we chose this idea.

What it does

Incident Claim Pack turns a short message like "My washing machine flooded the kitchen yesterday, help me with the insurance claim" into a complete claim pack, built only from the user's own NAS data:

1. Pins down the incident window from the user's words (or a transcribed voice memo) and confirms it in one line.
2. Camera events: queries the event timeline for that window and location (person, vehicle or motion events, depending on the incident type).
3. Damage photos and videos: finds media from the incident window and up to 7 days after it, asks the user which items really show the incident, then tags them and creates an album named after the claim ID. Both steps need the user's approval.
4. Invoices and warranty cards: finds documents for each affected item and extracts the item, model, serial number, purchase date, price, seller and warranty period. If two documents show different prices, it flags the difference and does not pick one silently.
5. Policy and warranty terms: answers the questions that matter, with short quotes and section numbers: Which clause covers this incident? Which exclusions could apply? What is the deductible? How many days are left to report? What evidence does the insurer require?
6. Claim summary: saves a structured summary with a timeline, evidence list, item table, policy findings and a missing-evidence checklist with a concrete next step for each gap.

The final reply tells the user the most urgent item, for example "23 days left to report, police report still missing".

How we built it

The Skill follows the Uliya Skill Development Standard and was built with the official uliya-skill-kit 0.7.7:

• We checked every step against the kit's capability catalog (9 CLIs, 84 Actions). Every step maps to a real Action, such as ug-surveillance event-timeline, ug-photo search-screen / create-album, ug-filemgr tag-add-file, ug-document extract / qa, and ug-knowledge save-text.
• SKILL.md is a pure orchestration Skill. It only delegates with delegate --target <agent> --instruction '...' to the five registered experts: security-monitor-agent, photo-agent, file-agent, document-agent and knowledge-agent. Each step passes only the real results of the previous step to the next one.
• It handles all five delegation states explicitly: completed, failed, cancelled, waiting_user and waiting_approval.
• Four Markdown references are linked directly from SKILL.md: workflow rules, the claim summary template, a test plan and a user guide.
• The package passes the kit's static checker with 0 errors and 0 warnings.

Challenges we ran into

• The web page and the kit did not say the same thing. The starter page lists "reconcile" under Documents. The kit catalog shows that ug-document reconcile only checks contract amounts against invoice amounts and cannot read insurance policies. We rebuilt the policy step around ug-document qa, which supports clause lookup with source citations.
• The first static check failed with 6 errors. A README in the package root is not allowed, the required sections were missing, and there were no machine-readable delegate commands. We restructured the Skill to match the checker's exact contract, including the capability-evidence format.
• Approvals. Tagging, album creation and saving text all require gateway or workflow approval. We redesigned the flow so that the user confirms the evidence first, and every write then waits for approval instead of happening silently.
• No invented data. Insurance is a domain where a guessed date or amount can cause real harm. Every missing value becomes "not found" plus a checklist item, and the Skill never says whether a claim will be paid.

Accomplishments that we're proud of

• One request now covers five NAS capability areas (Surveillance, Albums, Files, Documents and Knowledge) and produces an outcome people actually need.
• The Skill passes the official kit checker cleanly, and its safety rules are concrete: originals are never moved or deleted, nothing leaves the NAS, and card numbers are masked.
• The empty-result behaviour is honest. If nothing is found, the Skill says where it looked and stops. It never creates an empty album and never shows unrelated files.

What we learned

• Read the capability catalog, not the marketing description. Capability names can sound broader than what the Actions actually do.
• In an orchestration Skill, good design is mostly about boundaries: who does which step, what gets passed between steps, what happens when a step returns nothing, and where the user has to say yes.

What's next

• Validate the full flow on the assigned UGREEN AI NAS with sample data: a water-leak case, a driveway-theft case with camera events, and a "nothing found" edge case.
• Measure how reliably the local model quotes the right policy clauses, and tune the questions for long and non-English policies.
• Possible extensions: a reminder before the reporting deadline, and a retailer-ready warranty summary for appliance failures.

Third-party components: none. External services and data flows: none. All processing stays on the NAS.

Built With

• local-ai
• markdown
• ugos
• ugreen-ai-nas
• uliya
• uliya-skill-kit