scientific-slides – skill Tài liệu & Office cho agent AI

Scientific Slides

Overview

Scientific presentations are a critical medium for communicating research, sharing findings, and engaging with academic and professional audiences. This skill provides comprehensive guidance for creating effective scientific presentations, from structure and content development to visual design and delivery preparation.

Key Focus: Oral presentations for conferences, seminars, defenses, and professional talks.

CRITICAL DESIGN PHILOSOPHY: Scientific presentations should be VISUALLY ENGAGING and RESEARCH-BACKED. Avoid dry, text-heavy slides at all costs. Great scientific presentations combine:

• Compelling visuals: High-quality figures, images, diagrams (not just bullet points)
• Research context: Proper citations from research-lookup establishing credibility
• Minimal text: Bullet points as prompts, YOU provide the explanation verbally
• Professional design: Modern color schemes, strong visual hierarchy, generous white space
• Story-driven: Clear narrative arc, not just data dumps

Remember: Boring presentations = forgotten science. Make your slides visually memorable while maintaining scientific rigor through proper citations.

When to Use This Skill

This skill should be used when:

• Preparing conference presentations (5-20 minutes)
• Developing academic seminars (45-60 minutes)
• Creating thesis or dissertation defense presentations
• Designing grant pitch presentations
• Preparing journal club presentations
• Giving research talks at institutions or companies
• Teaching or tutorial presentations on scientific topics

Slide Generation with Nano Banana 2

The bundled image helper uses google/gemini-3.1-flash-image (Nano Banana 2),

with google/gemini-3.7-flash for visual review. API contracts, setup, and limitations

are in [references/script_reference.md](references/script_reference.md). Generated

images are drafts: verify every claim, label, citation, and scientific relationship.

There are two workflows depending on output format:

Image Workflow: PDF Slides

Generate conceptual or overview slides as complete images, then combine them into a PDF. Use editable PowerPoint or Beamer for quantitative results, exact equations, accessible text, or a user-supplied template.

How it works:

1. Plan the deck: Create a detailed plan for each slide (title, key points, visual elements)

2. Generate slides: Call Nano Banana 2 for each slide to create complete slide images

3. Combine to PDF: Assemble slide images into a single PDF presentation

Step 1: Plan Each Slide

Before generating, create a detailed plan for your presentation:

# Presentation Plan: Introduction to Machine Learning

## Slide 1: Title Slide
- Title: "Machine Learning: From Theory to Practice"
- Subtitle: "AI Conference 2025"
- Speaker: Dr. Jane Smith, University of XYZ
- Visual: Modern abstract neural network background

## Slide 2: Introduction
- Title: "Why Machine Learning Matters"
- Key points: Industry adoption, breakthrough applications, future potential
- Visual: Icons showing different ML applications (healthcare, finance, robotics)

## Slide 3: Core Concepts
- Title: "The Three Types of Learning"
- Content: Supervised, Unsupervised, Reinforcement
- Visual: Three-part diagram showing each type with examples

... (continue for all slides)

Step 2: Generate Each Slide

Use the generate_slide_image.py script to create each slide.

CRITICAL: Formatting Consistency Protocol

To ensure unified formatting across all slides in a presentation:

1. Define a Formatting Goal at the start of your presentation and include it in EVERY prompt:

• Color scheme (e.g., "dark blue background, white text, gold accents")
• Typography style (e.g., "bold sans-serif titles, clean body text")
• Visual style (e.g., "minimal, professional, corporate aesthetic")
• Layout approach (e.g., "generous white space, left-aligned content")

2. Always attach the previous slide when generating subsequent slides using --attach:

• This allows Nano Banana 2 to see and match the existing style
• Creates visual continuity throughout the deck
• Ensures consistent colors, fonts, and design language

3. Use the supplied presenter name and affiliation; omit them if unknown.

4. Include citations directly in the prompt for slides that reference research:

• Add citations in the prompt text so they appear on the generated slide
• Use format: "Include citation: (Author et al., Year)" or "Show reference: Author et al., Year"
• For multiple citations, list them all in the prompt
• Citations should appear in small text at the bottom of the slide or near relevant content

5. Preserve scientific figures as original assets:

• Locate the user's figures and underlying data before composing results slides.
• Embed the original chart unchanged in PowerPoint/Beamer, or regenerate it deterministically from the source data.
• --attach guides the image model; it can redraw points, axes, labels, and logos. It does not paste the original pixels or guarantee quantitative fidelity.
• Use attachments for style or conceptual layout drafts. For final results, add the original figure and verified text with a deterministic layout tool.

Example with formatting consistency, citations, and figure attachments:

# Title slide (first slide - establishes the style)
python scripts/generate_slide_image.py "Title slide for presentation: 'Machine Learning: From Theory to Practice'. Subtitle: '[Conference name and date]'. Speaker: '[Supplied presenter name]'. FORMATTING GOAL: Dark blue background (#1a237e), white text, gold accents (#ffc107), minimal design, sans-serif fonts, generous margins, no decorative elements." -o slides/01_title.png

# Content slide with citations (attach previous slide for consistency)
python scripts/generate_slide_image.py "Presentation slide titled 'Why Machine Learning Matters'. Three key points with simple icons: 1) Industry adoption, 2) Breakthrough applications, 3) Future potential. CITATIONS: Include at bottom in small text: (LeCun et al., 2015; Goodfellow et al., 2016). FORMATTING GOAL: Match attached slide style - dark blue background, white text, gold accents, minimal professional design, no visual clutter." -o slides/02_intro.png --attach slides/01_title.png

# Background slide with multiple citations
python scripts/generate_slide_image.py "Presentation slide titled 'Deep Learning Revolution'. Key milestones: ImageNet breakthrough (2012), transformer architecture (2017), GPT models (2018-present). CITATIONS: Show references at bottom: (Krizhevsky et al., 2012; Vaswani et al., 2017; Brown et al., 2020). FORMATTING GOAL: Match attached slide style exactly - same colors, fonts, minimal design." -o slides/03_background.png --attach slides/02_intro.png

# Results slides: insert figures/accuracy_chart.png as an original asset in
# PowerPoint or Beamer, then add verified findings and citations as text.
# The PowerPoint design guide below shows the explicit aspect-preserving geometry for this purpose.

# Conceptual methodology draft; provide the diagram file before running
python scripts/generate_slide_image.py "Presentation slide titled 'System Architecture'. Present the attached architecture diagram with brief explanatory bullet points: 1) Input processing, 2) Model inference, 3) Output generation. FORMATTING GOAL: Match attached slide style exactly." -o slides/04_architecture.png --attach slides/03_background.png --attach diagrams/system_architecture.png

Before creating results slides: find the original figure and source analysis, check labels,

units, sample sizes and uncertainty, then embed the figure without generative redrawing.

Numeric claims and placeholder references in these examples are illustrative, not evidence.

Prompt Template:

Include these elements in every prompt (customize as needed):

[Slide content description]
CITATIONS: Include at bottom: (Author1 et al., Year; Author2 et al., Year)
FORMATTING GOAL: [Background color], [text color], [accent color], minimal professional design, no decorative elements, consistent with attached slide style.

Accessibility handoff: Full-slide PNGs produce an image-only deck; their titles, equations, references, and data labels are not semantic text. Supply a transcript with slide numbers and descriptions of results figures. When the audience needs accessible slides, use the editable PowerPoint workflow with real title/text placeholders, reviewed alt text, and verified reading order; run PowerPoint's Accessibility Checker. Visual inspection alone cannot establish screen-reader usability. See [Microsoft's reading-order guidance](https://support.microsoft.com/en-us/powerpoint/make-slides-easier-to-read-by-using-the-reading-order-pane).

Step 3: Combine to PDF

# Combine all slides into a PDF presentation
python scripts/slides_to_pd

… (nội dung đầy đủ có trong gói tải về)