What inspired me
Before every exam I use AI to turn my lecture slides into notes and quizzes, but I could never fully trust the results. The notes sometimes skipped content, so I ended up reading them side by side with the original slides anyway. Other times they included parts that are never examined, like class activities, fun readings and further reading. The quizzes occasionally asked about things that were not in my slides at all.
In my courses, the exam tests the slides, and the slides follow the syllabus. So I wanted a study tool that treats my slides as the only source and can prove it.
What it does
Slide-to-Knowledge turns chapter PDFs stored on a UGREEN AI NAS into a trustworthy study pack:
- Files each PDF automatically into
Year-Semester / Subject / Chapter, by reading the title on page 1 and matching it against my saved subject list. It asks before moving anything. - Accounts for every page in an editable checklist: include, exclude (exercises, articles, learning outcomes, reading materials, class activities) or unsure. I make the final call.
- Writes topic-based notes using only the included pages, with a page citation on every point. Formulas are kept, diagrams are described with page references, and code goes into separate notes. It also flags any included page that the notes never cite.
- Runs quizzes (MCQ or structured, at theory, HOTS or mixed difficulty) that never go beyond the slides. Every question cites its page, MCQ answers explain every option, and structured answers are marked point by point with tips to improve.
How I built it
The Skill follows the Uliya Skill Development Standard. It is a SKILL.md that orchestrates three registered NAS experts: file-agent for folders and moving files, document-agent for reading, classifying and writing, and knowledge-agent for saving results. Each step uses only the real result of the previous one, and every move or save needs user approval and never overwrites existing files. I checked it with the official UGOS Skill Creation Tool, and it passes with no errors or warnings.
Challenges
- "Did it miss anything?" At first I could only tell which sections were left out, not which sentence or diagram. Switching to a page-by-page check solved this: every page is either cited in the notes or explicitly excluded with a reason.
- Working within the platform. Skills cannot include scripts, and there is no capability to copy images out of a PDF, so diagrams are described in text and linked back to their page instead of embedded.
- Finding my angle. Similar study Skills already exist. Comparing them showed me that they teach from the AI's own knowledge, while mine is strictly bound to the source.
What I learned
How NAS Skills are designed as orchestrations of expert agents rather than code, why approval and failure handling matter as much as the main flow, and that the hardest part of an AI study tool is not generating content but proving that the content is complete and in scope.
What's next
If selected as a finalist, I will test page classification accuracy, page citations and batch handling for long chapters on a real UGREEN AI NAS.
Built With
- document-agent
- file-agent
- knowledge-agent
- markdown
- skill.md
- ugos
- ugreen
- uliya
