Making Academic Presentations

Full reference: references/slide-generation.md

Tool: nanobanana (Gemini CLI extension)

Priority order (edit-first):

1. Has paper figure → nanobanana /edit to wrap into slide frame
2. Has existing slide → /edit to adapt
3. User-provided reference (e.g., from NotebookLM or PPTX the user made) → /edit to refine
4. Title slide from scratch → generate with academic style prompt
5. Content slide from scratch → generate with deck-style preamble

Key principle: prefer /edit on existing HQ paper figures over generating from scratch.

Deck style: create deck-style.md once per deck, prepend to all generate-from-scratch prompts. For /edit, style is inherited from the base image.

Example deck-style.md:

- Canvas: 1920x1080, white background
- Accent: #2563EB blue, text: #1e293b dark slate
- Clean sans-serif, flat design, no gradients/shadows
- Bottom bar: blue accent with white affiliation text

Stage 3: TTS Audio (optional)

Full reference: references/tts-engines.md Batch scripts: scripts/batch_tts_edge.py, scripts/batch_tts_kokoro.py

Output: one audio file per narrated slide

Engine Selection

Default: Use edge-tts unless user requests offline or premium quality.

Quick Start (edge-tts)

import edge_tts, asyncio

async def tts_slide(text, output, voice="en-US-AndrewNeural"):
    await edge_tts.Communicate(text, voice).save(output)

asyncio.run(tts_slide("Your slide text here", "slide_01.mp3"))

Voices: AndrewNeural (male, presenter), AriaNeural (female), GuyNeural (male, warm), JennyNeural (female, pro)

Stage 4: Video Assembly (optional)

Tool: ffmpeg Input: slide PNGs + audio files + optional demo recording

# Use symlink to avoid iCloud path spaces: ln -sfn "long path" /tmp/workdir

# Slide with audio:
ffmpeg -y -loop 1 -i slide.png -i audio.mp3 \
  -c:v libx264 -tune stillimage -pix_fmt yuv420p \
  -c:a aac -ar 44100 -ac 2 -shortest seg.mp4

# Silent slide (N seconds):
ffmpeg -y -loop 1 -i slide.png -f lavfi -i anullsrc=r=44100:cl=stereo \
  -c:v libx264 -tune stillimage -pix_fmt yuv420p \
  -c:a aac -ar 44100 -ac 2 -t N seg.mp4

# Concat (always re-encode, never -c copy):
printf "file 'seg1.mp4'\nfile 'seg2.mp4'\n..." > concat.txt
ffmpeg -y -f concat -safe 0 -i concat.txt \
  -c:v libx264 -pix_fmt yuv420p -c:a aac -ar 44100 -ac 2 final.mp4

All segments MUST share: 44100Hz sample rate, stereo, AAC codec.

PPTX Conversion (if needed)

Full reference: references/pptx-conversion.md

If starting from an existing PPTX, convert slides to PNG images first:

soffice --headless --convert-to pdf --outdir output/ presentation.pptx
pdftoppm -png -r 300 output/presentation.pdf output/slide

NotebookLM — Human Reference Only

The agent must NOT auto-invoke NotebookLM or use its outputs to drive slide/script decisions. The human owns the outline, visual arrangement, and deck direction.

When to recommend: only when the user says they're unsure what to put on slides or need inspiration.

Gotchas

• iCloud paths with spaces break ffmpeg — symlink to /tmp/
• Audio format mismatch breaks concat — always re-encode with -ar 44100 -ac 2
• ElevenLabs free tier — mp3_22050_32 only, 10k chars/month
• edge-tts needs internet — falls back to Kokoro if offline
• Kokoro WAV files are ~7x larger — convert to MP3 with ffmpeg before video assembly
• Kokoro first run downloads ~350MB model — ensure pip is in the venv
• /edit distorts figure — be more explicit: "Keep the original figure exactly as-is, only add framing"
• Style drift across slides — use /edit from base slide or prepend shared deck-style.md

Dependencies