Paper Deck — Visual Slide Director
Turn papers/knowledge content into slides that look professionally designed.
The core approach is not to rigidly arrange layouts using PPT objects, but rather:
- First understand the content, create a deck brief and page-by-page narrative.
- Clearly define for each page: "What do you want the audience to see, feel, and remember from this page?"
- Use an image generation model to produce 16:9 slide images.
- Synthesize into PPTX/PDF, and retain prompts as source files for revision.
When to Use
Suitable for:
- Paper group meetings, defenses, reading groups, technical sharing
- Visual presentations that "don't look like template PPT at first glance"
- Users who prefer high-quality image-based pages and prioritize overall visual impact and dissemination effect
- Need page-by-page revision: redo page N, change style, increase realism, reduce AI traces
Not suitable for:
- Need multiple people to edit each text box in PowerPoint with precision
- Large amounts of tables, financial statements, compliance materials
- Need to accurately replicate existing corporate PPT masters
If users require fully editable PPT, explain that V1 of this skill is raster-first; they can switch to regular PPTX tools, or use a hybrid solution of "image background + editable text layers".
Workflow
Step 1: Input Analysis
Accepts:
- arXiv / DOI / web links
- PDF paths
- Markdown / text / articles
- Existing outlines
- Reference images or reference PPT screenshots
For papers, prioritize reusing the reading method of
: read abstract, methods, experiments, charts, conclusions; search code repositories if necessary. The goal is not to write long articles, but to extract core narratives suitable for slides.
- Topic, audience, presentation scenario
- 1-sentence thesis of the paper/content
- 3-5 core points that must be clearly explained
- Recommended number of pages, recommended style, language
- Types of images to generate: cover, mechanism diagram, flowchart, data page, conclusion page, etc.
- Directly usable real materials: paper Figure/Table, PDF screenshots, user-provided screenshots, code screenshots, experimental curves
Step 2: Pre-Generation Confirmation
Confirmation is required by default; do not generate images directly unless the user explicitly says "generate directly/no need to confirm/follow defaults".
Limit questions to 3 or fewer when asking:
- Number of pages and purpose: group meeting / defense / public sharing / commercial presentation, how many pages are needed?
- Style: see
references/style-system.md
- Insert real materials: allow screenshots from PDF/paper charts, or user-provided screenshots/images? If allowed, specify which real materials are expected to be used on which pages.
Recommended script:
text
I suggest making 12 pages with the journal-minimal style: like Nature/IEEE paper figures + formal academic presentations, clear, restrained, and not flashy.
You can also switch to business-research for commercial research sharing, warm-notes for notebook style, or liquid-glass for Apple-style glass texture.
For this paper, I recommend inserting a partial screenshot of the original paper's method diagram on page 4, and a screenshot of experimental curves/tables on page 8, then doing design-based layout based on these real materials.
After confirmation, I will first generate outline.md and prompts for each page, then produce images page by page and synthesize into PPTX/PDF.
Step 3: Deck Brief
- (if any)
- : which pages use real charts/screenshots, sources and processing methods
Read
references/style-system.md
for style details as needed.
Read
references/source-visuals.md
for real material strategies as needed.
Step 4: Outline
Save as
. Each page uses a fixed structure:
markdown
## 01. Slide Title
- Role: cover / context / method / mechanism / evidence / result / takeaway
- Message: The only key point to be clearly explained on this page
- Visual: Main visual and composition of the page
- Text: Short text allowed on the page
- Evidence: Cited paper charts/formulas/experimental data/code locations
- Source visual: Whether to use real screenshots/paper charts; source, cropping range and placement
- Repair handle: Position description that can be referenced for subsequent revisions
Rules:
- Each page only carries one main point.
- Minimize text on the page; place complex explanations in speaker scripts or notes.
- For mechanism pages, prioritize drawing "input → processing → output" instead of abstract inspiration.
- For data pages, only include the 1-3 most convincing numbers.
- Real paper figures/screenshots are usually more credible than AI-generated ones; prioritize planning the placement of real materials when available.
- Do not overuse white space. The main visual, chart, or evidence area should usually occupy 60%-80% of the page, except for covers or section pages.
- For decks with more than 8 pages, must have rhythm changes: alternate between cover, problem, method, mechanism, evidence, conclusion.
Step 5: Prompt Files
Must write a prompt file for each page before calling any image generation tool.
Path:
text
paper-deck/{topic-slug}/
├── analysis.md
├── deck-brief.md
├── outline.md
├── prompts/
│ ├── 01-slide-cover.md
│ ├── 02-slide-context.md
│ └── ...
├── images/
│ ├── 01-slide-cover.png
│ ├── 02-slide-context.png
│ └── ...
├── {topic-slug}.pptx
└── {topic-slug}.pdf
Read
references/prompt-template.md
for prompt writing guidelines.
Hard rules:
- Prompt must explicitly specify 16:9.
- Prompt must clearly state style, composition, text language, and text quantity limits.
- Do not let the model generate page numbers, logos, watermarks, or PPT frames.
- If precise text is needed, minimize text within the image; a hybrid text layer can be added later.
- If real materials are used on this page, the prompt must explain how the materials are integrated into the page: embedded, cropped, carried by glass panels, annotated, magnified, rather than letting the model redraw facts from scratch.
Step 6: Generate Images
Image backend options:
- Use built-in first in Codex environment.
- If the user specifies backends like , Gemini, OpenAI, Seedream, follow the user's choice.
- If no available image generation backend is present, stop and inform the user that a raster image backend is required.
Generation strategy:
- First generate page 1 as the style anchor.
- For subsequent pages, if the backend supports reference images, use page 1 as the style reference to reduce style drift.
- Check thumbnails every 3-4 pages; if style drift is found, revise the prompt before continuing.
- Save failed pages; do not overwrite successful ones.
Step 7: Synthesize PPTX/PDF
After generating images, run:
bash
python3 <SKILL_ROOT>/scripts/merge_deck.py paper-deck/{topic-slug}
The script will read
and output a corresponding
and
. Each image fills a full 16:9 page.
Step 8: Quality Check
Check against
references/quality-gate.md
before delivery:
- Does it look like a real design work at first glance, rather than a template stack?
- Does each page only carry one main point?
- Is there excessive meaningless white space; does key content occupy enough of the page?
- Are real material pages clearly recorded with source, page number/figure number and placement?
- Is the style consistent?
- Is the text in the image clear, free of typos, and not gibberish?
- Are common AI issues present: fake UI, fake logos, garbled labels, excessive cyberpunk style, plastic 3D, meaningless decorations?
- Can the PPTX/PDF be opened, and is the number of pages correct?
Step 9: Revision
Always modify source files first when revising:
| User says | Action |
|---|
| "Make page 5 more academic" | Revise , keep the old image, generate a new one |
| "Unify the texture to match page 1" | Add the style anchor of page 1 to relevant prompts |
| "Too much text on page 7" | Modify the Text section in the outline, then revise the prompt |
| "Only redo the background, don't change the content" | Keep Message/Text in the prompt, rewrite the Visual section |
| "Add a new page for mechanism details" | Update the outline, add a new prompt, generate the image, rerun the synthesis script |
Do not use programs to alter text on generated images. If text is incorrect, revise the prompt and regenerate, or switch to the hybrid text layer solution.
Reference Files
references/style-system.md
- : Common page roles and compositions
references/source-visuals.md
references/prompt-template.md
references/quality-gate.md