Inspiration
Most households and small offices keep a folder of contracts: insurance, mobile, gym, parking, software. Many renew automatically unless notice arrives in time, often three months before the term ends, and a missed date can cost another six or twelve months. The notice rules sit in the small print, sometimes in four languages.
What it does
Cancel Window is a Uliya Skill for the UGREEN AI NAS. Ask "Which contracts must I cancel or review in the next 6 months?" and it reads the contracts in the folder you name, extracts the term end, notice period and renewal rule with verbatim quotes, and computes the last day to give notice. Every date comes with its clause and a one-line calculation such as "31 May 2027 minus 3 months = 28 Feb 2027". The digest sorts agreements into due, window passed, ends automatically, outside the horizon, open-ended and check manually. Missing, unreadable, conflicting or law-defined terms become "check manually" instead of a guess. After approval it saves the digest and cancellation letter drafts as new files and can tag the contracts. It reads contract terms only. Not legal advice; statutory cancellation rights may differ.
How we built it
The Skill is Markdown only: SKILL.md plus four references (date rules, digest format, calendar export, disclosure). It delegates to UGREEN's file-agent (ug-filesystem, ug-filemgr), document-agent (ug-document) and knowledge-agent (ug-knowledge), passing only real upstream results. Writes happen only after approval and never overwrite. A synthetic set of 20 fictional contracts plus one non-contract in English, German, French and Italian has Python-checked expected answers.
Challenges
Date arithmetic is full of traps: 31 May minus 3 months, notice "to the end of a month", weeks instead of months, notice that changes after renewal, open-ended agreements and contracts that defer to the law.
What we learned
An offline desk dry run (no NAS, experts simulated from the PDF text by the same assistant that wrote the Skill) scored version 1 at 70 to 76 of 84 fields. It showed that open-ended agreements could vanish from the digest. Version 2 fixes this and scores 84 of 84, in-sample. These are simulations, not NAS results.
What's next
Validate on the assigned NAS in Round 2: extraction accuracy, letters, approvals, reruns and an empty folder.
AI use and authorship
Grok Bot, an AI assistant, generated the idea, wrote the Skill and references, built the synthetic test set and drafted the proposal. Marco Costa Padilla is the entrant: he approved the idea and the submission. In Round 2 the assistant operates the test session under his login with his approval. UGREEN's check-skill.mjs (0 errors, 0 warnings) was used for validation.
Log in or sign up for Devpost to join the conversation.