Inspiration
Every household has a drawer of fading thermal receipts, a phone full of photographed bills and a folder of PDF order confirmations. When an appliance breaks, the question is always the same: is it still under warranty, and where is the receipt? In an EU-wide European Commission study, only 35% of consumers knew the legal guarantee period in their own country, and 17% of unresolved product problems were never claimed because people wrongly believed the guarantee had expired. A NAS already holds all of that evidence; it just does not know what it means.
What it does
Warranty Proof Pack is a Uliya Skill with two modes.
- Register: "Which of my purchases are still under warranty?" The Skill finds proofs of purchase in the folders and photo library you allow, reads merchant, item, date and price, applies the EU 24-month legal guarantee (or a longer commercial warranty printed on the receipt), and saves a dated register in which each item is marked active, expiring soon, expired or date not found.
- Claim pack: "My espresso machine stopped heating, prepare a claim." The Skill finds the purchase, proposes product photos, fingerprints every evidence file with SHA-256, copies them into a new claim folder and drafts a letter to the seller from verified facts only. You review it and send it yourself.
Nothing leaves the NAS, nothing is deleted or overwritten, and every write waits for your approval.
How we built it
The Skill is a pure orchestration Skill that follows the Uliya Skill Development Standard. The parent Skill only delegates to the registered expert agents:
- file-agent: find candidate documents, SHA-256, folders, copies, tags;
- photo-agent: OCR-based receipt search, receipt reading, product photos;
- document-agent: receipt and invoice extraction, register lookup, letter drafting;
- knowledge-agent: saving the register, the letter and the fingerprint list.
Each step passes only the real result of the previous one forward. All five delegation states (completed, failed, cancelled, waiting_user, waiting_approval) are handled explicitly. Every capability was checked against the official capability catalog in uliya-skill-kit 0.7.3.
Challenges
- Never inventing data. A faded receipt must produce an empty field and a clear "could not be read", not a plausible guess.
- Merging duplicates. One purchase often exists both as a photo and as a PDF.
- Explaining legal time limits honestly. Every row shows the rule that was applied, and the register says it is guidance, not legal advice.
What we learned
The value of an AI NAS is not one more chat box but joining libraries the user already has (files and photos) and turning them into decisions with deadlines and evidence.
What's next
On the NAS validation day: measure extraction accuracy on a fictional set of 22 proofs of purchase, including faded receipts, and confirm the approval prompts on copy and save. Then add more country rules.
AI tools: Claude Code assisted the drafting under the author's direction. The full disclosure is in references/disclosure.md inside the Skill package.
Built With
- claude-code
- markdown
- ugos
- uliya
