Exam Radar

Turns a course folder of notes and past-question photos on a UGREEN AI NAS into a revision pack: which topics keep getting examined, what your notes miss, and cited practice answers.

Comment

Inspiration

Before every exam in my department, past questions travel around as phone photos in group chats, and lecture notes end up split across PDFs, slides and pictures of handwritten pages. The question everyone asks is the same: which topics come up every year, and do my notes actually cover them? I have built past-question study tools before, and this felt like the right job for a NAS that already holds all of those files.

What it does

Exam Radar reads one course folder on a UGREEN AI NAS and saves a revision pack back into it:

  • a topic map of the past papers: every examined topic, how often it appears and in which papers
  • a coverage check of the student's notes against those topics: covered, thin or missing
  • a revision summary for covered topics, with the file and page or photo behind every point
  • practice questions in the style of the past papers, answered only from the student's notes. Topics the notes miss get questions but no invented answers, under "Not in your notes"

How I built it

The Skill is a single orchestration SKILL.md built with the UGREEN Uliya Skill Kit 0.7.3. It delegates six ordered steps to the NAS's registered experts:

  1. file-agent (ug-filesystem) lists the course files
  2. photo-agent (ug-photo vision-ocr) reads photos of notes and past papers
  3. document-agent (ug-document extract/qa) maps the examined topics
  4. document-agent checks the notes against each topic
  5. document-agent (ug-document writing) drafts the pack
  6. knowledge-agent (ug-knowledge save-text) saves one new Markdown file after the user approves

Everything is read-only until the last step, every delegation state (completed, failed, cancelled, waiting_user, waiting_approval) is handled, and unreadable photos are listed instead of guessed. The package passes the kit's static checker with no errors or warnings.

Challenges

The kit's documentation and its checker disagree on a few format details (the capability evidence heading and arrow, and the workflow heading), so I worked through the checker diagnostics until it passed cleanly. The harder design problem was stopping the pack from inventing answers: every step passes only the real result of the step before, and answers are allowed only where the notes support them.

What's next

Testing on a real NAS: OCR quality on messy handwriting, how well topics group across different years' papers, and the approved save.

AI disclosure

Claude Code (Anthropic) drafted the SKILL.md workflow, this text and the proposal from my chosen idea, and ran the kit's static checker.

Built With

Share this project:

Updates

Submission history