Inspiration

I study at TUM Campus Heilbronn. My courses are in English, which is my second language, so I usually record lectures. But I almost never listen to them again, it'll cost a lot of time. Before an exam nobody has time for hours of audio. The parts I actually need are the ones that are not on the slides: an extra example, a spoken correction like "slide 8 is wrong", or "this will be on the exam".

What it does

Lecture Delta Ledger is a Skill for UGREEN AI NAS. It compares one lecture recording with the PDF slides for that lecture and reports only the differences:

  • Spoken-only: said in the lecture, not on the slides
  • Slide-only: on the slides, not found in the recording
  • Possible conflict: the lecturer said something different from the slide
  • Exam cue: the lecturer explicitly mentioned an exam, quiz or homework

Every item comes with a quote, its position in the recording and the slide page. Exam cues from all weeks go into one ledger per course, so before the exam I can ask: what did the lecturer say will be tested?

How I built it

The Skill is Markdown only and follows the Uliya Skill Development Standard. It uses three registered agents:

  1. file-agent finds the recording, the slides and earlier reports in the course folder.
  2. document-agent checks that the PDF has real text and reads it page by page.
  3. knowledge-agent transcribes the recording on the NAS.
  4. document-agent compares the transcript with the slides.
  5. After I confirm, the report and the ledger are saved as Markdown files.

I used AI tools while building it: Claude suggested the initial idea, reviewed and revised the workflow, and helped with wording. Google Antigravity (Gemini) drafted the first version of the Skill files.

Challenges

  • The comparison step only reads real files on the NAS, so the transcript has to be saved first. I added a confirmation step and a fallback for when saving is not possible.
  • The Skill must never invent evidence. If something is missing from the slides, it reports which pages it checked instead of making up a page number. If the recording and the slides are about different topics, it stops and asks.

What I learned

How to design a workflow around what the NAS agents really return, and how to write down clearly what happens when a step fails or the user says no.

What's next

The Skill passes the official checker, but it has not run on a real NAS yet. If I become a Finalist, I will test it with a recording I make myself and slides with planted differences, then check that re-running a week does not create duplicates.

Built With

  • markdown
  • ugos
  • ugreen-ai-nas
  • uliya-skill-kit
Share this project:

Updates

Submission history