Lecture Slides
Turns a lecturer's rough drafts into four deliverables that stay in sync:
| Deliverable | Where |
|---|---|
Slide deck (Slides artifact, one <aside> of notes per slide) |
claude.ai artifact |
N강_최종대본.txt — final script, N페이지 + 코드 N sections |
lecture folder |
N강_일레븐랩스_PPT.txt, N강_일레븐랩스_코드.txt — TTS scripts |
lecture folder |
NN_topic/NN_topic.ipynb + README row |
course GitHub repo |
The look, the slide recipes, the script rules and the notebook rules are in the reference files. Read the one you need before starting that step:
reference/style.md— palette, type scale, slide archetypes, layout budgetreference/slides-howto.md— artifact mechanics, diagrams, images, revisingreference/script.md— polishing the draft, final script format, ElevenLabs rulesreference/notebook.md— notebook conventions, local run, Colab check, GitHubtemplates/*.html— copy-ready slide files for each archetype
Workflow
- Read everything first. The script draft (
N강 대본 초안.txt), the PPT draft (export to PNG to look at it —reference/slides-howto.md), any images in the folder orbackup/, and the notebook. Drafts may be mojibake in the first half and clean UTF-8 later; find the readable copy. - Ask only what changes the work (one AskUserQuestion, ≤3 questions):
- notebook folder/file name (
NN_topic/NN_topic.ipynb, recommend one), - the user's own figures: redraw in deck style, or insert as images (if the user drew a figure carefully, inserting it is usually right — see "Lessons"),
- anything the draft leaves open (which experiments, who writes a missing script section). Everything else: decide, state the assumption in one line, go.
- notebook folder/file name (
- Notebook (if any) →
reference/notebook.md: rename, format like the previous lecture, fix wrong comments, run locally, push, README row. Get the numbers the script will quote from the run. - Deck →
reference/slides-howto.md+reference/style.md: one idea per slide, notes = the polished script for that page, citations for paper figures, a closing "코랩에서 직접 실행" slide with the notebook URL when there is a notebook. - Scripts →
reference/script.md: the review draft first (N강_대본.txtorN강_코드대본.txt) when the user wants to review; the final script and ElevenLabs files once they have reviewed. - Report: what changed, the numbers used and where they came from, and a short "확인해 주세요" list (factual fixes you made to their wording, pronunciations you guessed, anything not verified). Never claim a render or a Colab run you did not do.
Rules that came from the user
- Commit/push only when asked; the user asks explicitly ("올리고 확인", "깃허브에 올려줘"). Never delete their files; keep originals beside the renamed copy.
- Their edits win. The deck is edited in the browser and sometimes in a downloaded PPTX. Before touching a slide, read its current version from the artifact; change only what was asked; keep their widths, wording, images.
- Numbers follow Colab, because the video is recorded on Colab. If you only ran locally, say so and round ("약 0.29").
- The last code scene ends with the next-lecture line ("다음 시간에는 … 살펴보겠습니다."), then "감사합니다." Ask or leave a placeholder if the next topic is unknown.
- Korean throughout, polite 합니다체 in scripts; slides use short noun-phrase or 다/음 endings.
Lessons
- A redrawn diagram can lose what the user's figure showed (lecture 8: the residual-connection backprop flow). When the user supplies a careful figure, insert it (upload as asset) and put short takeaway cards beside it.
- Formula images the user pastes (PNG of LaTeX) go in a white card
(
background:#FFFFFF) at native size — scaling up blurs them. - When the user's wording is technically wrong (e.g. "squaring the gradient = second derivative", "Wide ↔ Deep inputs swapped"), fix it in the script and list it under "확인해 주세요"; do not silently keep or silently change.
- Verify every quoted number against the notebook output; compute any derived claim (e.g. "baseline loss equals predicting the mean") before saying it.