Your slides, animated.
Transform any presentation into a cinematic experience with GPT Image slides and Veo 3.1 video transitions.
Every deck is the same: title, bullets, stock photo. PowerPoint transitions feel dated. No one remembers your slides the next day.
Professional-looking decks require designers, hours of manual layout work, and expensive tools. Most teams skip it entirely.
Cinematic transitions between slides would transform presentations, but creating them requires professional motion design and video editing.
The gap between "a deck someone made in 30 minutes" and "a deck that makes the audience feel something" is a professional design team and a week of production time.
Director Deck takes a text prompt, a PPTX file, or an HTML presentation and produces a polished deck with AI-generated slide images and cinematic video transitions — reviewed at every stage.
Describe your deck in plain language. The Ghost Deck Writer crafts 8–12 slides with assertion titles, a storytelling framework, and a full visual identity.
Drop in a PowerPoint file. Content, speaker notes, theme colors, and fonts are extracted verbatim. Slide count is preserved exactly.
Pass a URL or local file. Playwright screenshots each slide at 1536×864. DOM text and CSS design tokens are extracted automatically.
A staged recipe with human approval at every creative checkpoint:
Creates slide_deck.json and DESIGN.md — slide content, storytelling framework, visual identity, and pre-written Veo transition prompts.
Builds wireframe HTML and PPTX previews. Human reviews layout, content flow, and visual identity before image generation begins.
GPT Image 2 renders each slide as a complete 2560×1440 image — atmosphere, typography, layout in one shot. Human reviews before transitions.
Veo 3.1 generates 8-second cinematic video clips between each pair of slides. Post-processed with ease-in-out retiming via ffmpeg.
Assembles final_deck.pptx with interstitial video slides and deck_live.html for browser playback. Patches four python-pptx bugs automatically.
Content authoring — titles, bullets, speaker notes, DESIGN.md
Wireframe HTML + PPTX previews from slide_deck.json
GPT Image 2 renders — 2 images per slide at 2560×1440
Veo 3.1 cinematic clips — Playwright keyframes to video
Final assembly — PPTX with interstitial videos + HTML player
All five agents share slide_deck.json as a Pydantic v2 data contract. Each agent enriches the JSON at its stage — assets, transition paths, and timing accumulate as the pipeline progresses.
Every image must be 16:9 or transitions show a visible edge jump at 1920×1080. GPT Image is called at 2560×1440; Pillow crops to 1536×864 for Veo input. Zero tolerance for aspect ratio drift.
The words presentation, slide, deck, infographic, PowerPoint, chart, and data visualization are banned from all Veo prompts. They cause Veo to hallucinate slide content in intermediate frames.
Veo provides 8 seconds of source material. video_processor.py holds the first and last 0.4s at natural speed, compresses the middle for smooth ease-in-out — real Veo frames, no interpolation artifacts.
pptx_fixer.py patches four python-pptx bugs: broken click-to-play hlinkClick, missing autoplay, shared poster frames, and absent auto-advance timing.
A 10-slide deck produces 19 PPTX slides: 10 content + 9 video interstitials. The only valid OOXML approach — <p:transition> does not support embedded video.
Transition duration is an editorial decision. Contemplative decks get 4–6s transitions. Business decks get 2–3.5s. Emotional pivots add +1.5–2s automatically.
All text must fall within x:120–1416, y:130–734 on the 1536×864 canvas. Content outside the safe zone gets cut off at various display sizes.
A complete PowerPoint file with full-bleed AI-rendered content slides interleaved with auto-playing video transition slides. Auto-advances in Presenter mode. Open in PowerPoint and present — no additional setup needed.
An interactive HTML player with the same slide + transition experience in the browser. Keyboard and touch navigation. No server required — open the file directly.
Three ways to invoke the pipeline:
Each gate pauses for human review. Edit slide content, visual identity, or individual images at any checkpoint. Approve with revision instructions like "redo slides 3,5 — text too small on 3, wrong colors on 5" and only those slides regenerate.
director_deck/agents/Director Deck proves a pattern: take a creative medium that was locked behind professional tools, decompose it into a pipeline of specialized AI agents with human checkpoints, and make the result accessible to anyone who can describe what they want.
Inspired by Dan Shapiro's DirectorDeck process — an approach to creating AI-animated presentations that Director Deck packages into a reproducible, reviewable, fully automated Amplifier recipe.
All data in this presentation comes from the kenotron-ms/amplifier-bundle-director-deck repository, verified May 2026.
git log --oneline --since="60 days ago" --no-merges | wc -l → 31 commitsgit log --format="%an" | sort -u → Ken Chau (kenotron-ms)wc -l director_deck/*.py → 1,951 lines across 7 moduleswc -l tests/*.py → 707 lines across 8 test filesagents/ (ghost-deck-writer, slide-architect, visual-director, transition-director, deck-stitcher)director-deck.yaml (v0.5.0, 4 stages, 3 approval gates) and director-deck-continue.yamlteam_knowledge(operation="search", query="director deck") → 5 indexed capabilities
Technical claims (16:9 enforcement, Veo banned words, PPTX bug fixes, interstitial slide approach, transition easing) verified directly from README.md, recipes/director-deck.yaml changelog, and source modules (video_processor.py, pptx_fixer.py, schema.py).
No metrics were fabricated. Where the deck describes pipeline behavior, the source was the recipe YAML (891 lines with inline agent prompts) and the README (230 lines).