Fix-It Shelf

Inspiration

When the washer shows an error code, the answer is usually in its own manual - somewhere among thousands of scanned files on the NAS - and the next question is whether the warranty still covers a repair. The model number is on a sticker behind the machine, the receipt is in another folder, and general AI tools answer from the web for a generic model, not for yours.

What it does

Fix-It Shelf reads the folders you point it at and builds one card per thing you own: manual, receipt, warranty card and sticker photo, grouped by model and serial. Then:

• "The dishwasher shows E24" - you get the steps from your own manual, with the file and page they came from, and whether the warranty still covers a repair.
• "What goes out of warranty in the next 30 days?" - a list from your own receipts.
• Anything it is unsure of becomes a question to you ("Does this receipt belong to the dishwasher bought in 2021 or the one bought in 2024?").

It never looks anything up on the web, never guesses from another model's manual, and never moves, renames or deletes your files. Nothing leaves the NAS.

How we built it

An orchestrator Skill following the Uliya Skill Development Standard, with three expert roles delegated through the kit's agents: an Indexer (files find/browse and document extraction of model, serial, date and warranty length), a Matcher (cards per model and serial, tolerant of O/0 and I/1 sticker misreads, every uncertain link turned into a user question) and an Answerer (searches one item's manual only, cites the page, stops when the code is not there). The shelf is saved as one knowledge file where the user approves.

Challenges

• The official kit checker was strict about structure: section names, real delegate lines per step, and handling of all five delegation states. The package now passes uliya-skill-kit 0.7.3 with 0 errors and 0 warnings.
• Twin appliances: two identical models bought years apart must stay two cards, split by serial and purchase date.

What we learned

Most of the value is in refusing to guess: an answer that cites your own manual's page is worth more than a fluent answer about a similar model.

What's next

Round 2: real-scan OCR on a UGREEN AI NAS, then the two core tests (30 mixed files for 6 appliances on the right cards; 10 error-code questions answered from the right page) and the twin-model edge case.

AI disclosure

Designed and written with an AI assistant (Claude, by Anthropic), which drafted the Skill instructions and the proposal; offline logic checks were run on a synthetic 60-document shelf.

Built With

• claude
• markdown
• ugos
• uliya-skill-kit