Generate Writeup

Generate annotated write-up from a presentation

You are generating a blog-style write-up from a presentation. The user will provide a presentation folder path (e.g., presentations/my-talk). That folder must contain a presentation.md file.

presentation.md format

# Talk title

- **Date:** 2026-01-15
- **Video:** https://www.youtube.com/watch?v=VIDEO_ID
- **Slides:** slides.pdf
- **Transcript:** transcript.txt

## Notes

Extra notes here

Required fields: Video and Slides. All others are optional. Slides can be a local PDF path, PDF URL, PPTX URL, OneDrive link, or RevealJS URL. Video can be a YouTube URL or local MP4 path.

Pipeline

Follow these steps in order. Cache all intermediate outputs in <presentation_folder>/outputs/ so subsequent runs skip completed steps.

Step 1: Resolve slides

• If Slides in presentation.md is a URL, fetch and convert to PDF:

  uv run .agents/skills/fetch-slides/fetch_slides.py <url> <presentation_folder>/outputs

  This produces slides.pdf (and slides_content.md for RevealJS URLs).
• If slides is a local file path (relative to the presentation folder), use it directly.
• Skip if <presentation_folder>/outputs/slides.pdf already exists.

Step 2: Convert slides to images

Split the PDF into individual PNGs:

uv run .agents/skills/convert-slides-to-images/convert_slides_to_images.py <pdf_path> <presentation_folder>/outputs/slide_images

Skip if <presentation_folder>/outputs/slide_images/ already contains slide_*.png files.

Step 3: Extract transcript

• If Transcript is specified in presentation.md, read from that path (relative to presentation folder).
• If <presentation_folder>/outputs/transcript.txt exists, use the cached version.
• Otherwise, extract from YouTube:

  uv run .agents/skills/extract-transcript/extract_transcript.py <youtube_url> <presentation_folder>/outputs/transcript.txt

Step 4: Generate video chapters

If <presentation_folder>/outputs/chapters.txt exists, use the cached version. Otherwise:

1. Read the transcript from Step 3.
2. Write a brief summary paragraph (2-3 sentences) describing what the video is about.
3. Create a list of timestamped chapters in "MM:SS - Chapter Title" format covering the main topics.
4. Save to <presentation_folder>/outputs/chapters.txt.

Step 5: Extract slide text

Extract the text content of each PDF page to provide ground truth for what each slide actually says. This prevents misidentifying embedded screenshots or demo captures as actual slide content.

uv run .agents/skills/extract-slide-text/extract_slide_text.py <pdf_path> <presentation_folder>/outputs/slide_ascii.md <presentation_folder>/outputs/slide_images

Skip if <presentation_folder>/outputs/slide_ascii.md already exists.

Step 6: Outline slides

If <presentation_folder>/outputs/outline.txt exists, use the cached version. Otherwise:

1. Find all slide_*.png files in <presentation_folder>/outputs/slide_images/, sorted numerically.
2. Read <presentation_folder>/outputs/slide_ascii.md (from Step 5) as ground truth for each slide's text content.
3. Look at each slide image and write a one-sentence summary. Base the summary primarily on the extracted text from slide_ascii.md, using the image only for visual context (diagrams, layout, screenshots).
4. For large presentations (50+ slides), work in batches of 50.
5. Save to <presentation_folder>/outputs/outline.txt.

Step 7: Generate the annotated write-up

Gather all context, then generate the write-up following the rules below.

Context to gather:

• TRANSCRIPT — Full transcript text (Step 3)
• VIDEO_CHAPTERS — Chapter summary (Step 4)
• SLIDE_TEXT — Extracted text per slide from slide_ascii.md (Step 5). Use as ground truth for what each slide contains.
• SLIDE_OUTLINE — Numbered slide outline (Step 6)
• VIDEO_SOURCE — YouTube URL or MP4 path from presentation.md
• IS_LOCAL_VIDEO — true if VIDEO_SOURCE is a local MP4 file
• SLIDES_HTML_CONTENT — Contents of slides_content.md if it exists (RevealJS only)
• Check if SLIDES_HTML_CONTENT contains "SECTION HEADING" markers → HAS_SECTION_HEADINGS

Write-up generation prompt:

Create an annotated blog post that explains the content of each slide.

STRUCTURE REQUIREMENTS:

1. Start with a level-1 heading (#) containing the talk title.
2. Follow with an overview paragraph introducing the talk.
3. Do NOT include a table of contents — it will be generated automatically.

If HAS_SECTION_HEADINGS is true (RevealJS slides with section markers):

• Use level-2 headings (##) for SECTION HEADING slides — these are section dividers.
• Use level-3 headings (###) for regular slides within each section.

If HAS_SECTION_HEADINGS is false:

• Use level-2 headings (##) with a descriptive title for each slide's content.