AI Presentation Workflow

Create professional presentations: Content → Design → Build → Assembly → Polish.

Step 0: Choose Workflow Settings

At the start of every presentation task, ask the user TWO choices:

0-A. Collaboration Mode

Mode	Description	Checkpoints
Full Auto	Minimal interaction. Confirm topic only, deliver final PPTX.	1 checkpoint
Guided (recommended)	Confirm outline, pick design, preview before assembly.	3 checkpoints
Collaborative	Review every slide, approve every illustration, full control.	Per-slide

If the user doesn't specify, default to Guided mode.

0-B. Assembly Method

Method	How it works	Best for
Editable HTML (Path A)	HTML slides + selective AI illustrations → html2pptx → editable PPTX	Need to edit text later, precise layout, corporate decks
Full AI Visual (Path B)	Every slide as a complete AI-generated image → create_slides.py → image PPTX	Maximum visual impact, artistic presentations, quick drafts

Trade-offs:

	Path A: Editable HTML	Path B: Full AI Visual
Text	Editable in PPT	Baked into image (not editable)
Visual quality	Good with illustrations	Excellent — cohesive design
Layout control	Pixel-precise	AI-interpreted
File size	Smaller (~5-25MB)	Larger (~30-80MB)
Chinese text	Perfect (font rendering)	Usually good (AI may occasionally misrender)
Speed	Faster (HTML creation)	Slower (image generation per slide)

If the user doesn't specify, default to Path A (Editable HTML).

Step 1: Content Structuring

Turn raw material into a slide-by-slide outline.

Per slide, define:

Title — a complete assertion sentence (not a topic word)
Key points — 3-4 maximum
Visual type — illustration / chart / diagram / icon / quote
Path A: Illustration needed? — Yes/No. If yes, one-line description.
Path B: Visual scene description — one paragraph describing the complete slide visual (layout + imagery + mood).

Assertion-Evidence rule:

Bad title	Good title
Q3 Sales	Q3销售增长23%，新用户是主要驱动力
Methodology	我们通过双盲实验验证了这个结论

语言规则：slide内容一律用中文，仅保留必要的英文术语（人名、品牌名、技术专有名词）。 Section label（如 INSIGHT、TAKEAWAY）可用英文作为设计元素。

✅ Checkpoint 1 (Guided + Collaborative)

Present the outline as a table:

Path A:

| # | Title (assertion) | Key Points | Visual Type | Illustration? |
|---|-------------------|------------|-------------|---------------|
| 1 | Cover: ... | — | Decorative | Yes: ... |
| 2 | ... | 1. ... 2. ... | Chart | No |
| 3 | ... | 1. ... 2. ... | Illustration | Yes: ... |

Path B:

| # | Title (assertion) | Key Points | Visual Scene Description |
|---|-------------------|------------|--------------------------|
| 1 | Cover: ... | — | Dark gradient bg, large title centered, abstract network nodes |
| 2 | ... | 1. ... 2. ... | Split layout: text left, bar chart right, clean white bg |
| 3 | ... | 1. ... 2. ... | Full illustration: person at crossroads with floating clocks |

Ask the user:

Approve / adjust slide count
Path A: Approve / adjust which slides get illustrations
Path B: Approve / adjust visual scene descriptions
Any content to add or remove

Step 2: Design System

Present 3 design system options for the user to choose from. Each is a complete visual language, not just a color palette.

CRITICAL: A design system is NOT just colors. It defines visual philosophy, typography ratios, composition rules, and emotional intent. This is the difference between "boring PPT" and "magazine-quality deck."

🗣️ Style Discussion (Optional, if user wants to explore)

If the user says things like:

"我想要XX风格"（田中一光、瑞士国际主义、包豪斯、蒙德里安...）
"我不确定想要什么风格"
"能给我看看不同风格的例子吗"

Then consult the design movements reference: references/design-movements.md — 设计运动与风格参考库

This file maps classic design movements (Neo-Brutalism, Swiss Style, Bauhaus, etc.) to our AI-ready style presets. Use it to:

Translate user's aesthetic language into actionable prompts
Build shared vocabulary ("这个方向偏田中一光" vs "那个偏构成主义")
Reference when designing new custom styles from scratch

After discussing movements, proceed to recommend 3 concrete presets below.

Design System Presets

⚠️ CRITICAL INSIGHT: 插画/漫画类风格的AI生成效果远好于「专业极简」类风格。 漫画/插画风格有明确的视觉语言（线条、角色、色块），AI可以充分发挥；极简风格（暗色底+发光文字+大量留白）缺乏视觉元素，生成出来「空」且「平」。

Pick 3 that match the topic/mood. Use the topic recommendation table below, then present each with its full description.

按主题自动推荐（优先从此表选）：

主题类型	第一推荐	第二推荐	第三推荐
品牌/产品介绍	Snoopy温暖漫画	Neo-Pop新波普	浮世绘/敦煌（东方品牌）
教育/培训	Neo-Brutalism	学習漫画	Snoopy温暖漫画
技术分享	xkcd白板	Neo-Brutalism	Ligne Claire
数据报告	Ligne Claire	苏联构成主义	Neo-Brutalism
年轻受众	Neo-Pop	像素画	孔版印刷
创意/艺术	达达拼贴	孔版印刷	The Oatmeal
国风/东方	敦煌壁画	浮世绘	温暖叙事
正式商务	Neo-Brutalism	Ligne Claire	温暖叙事
产品发布/keynote	苏联构成主义	Neo-Pop	Neo-Brutalism
内部分享	Neo-Brutalism	The Oatmeal	xkcd白板

完整18种风格详细参考： references/proven-styles-gallery.md 风格样例图片： assets/style-samples/ 目录

第一梯队（强烈推荐，效果极好）：

1. Warm Comic Strip — Snoopy温暖漫画风

Philosophy: Peanuts漫画的温暖与哲理感——简单的角色说着深刻的话，日常场景中蕴含人生智慧
Visual world: 圆头小孩、小狗、小鸟组成一个温暖的小世界。背景极简（草地、天空、狗屋、树）。色调像泛黄的报纸漫画
Reference: "Like a Peanuts comic strip — warm, philosophical, charming"
Style guide: references/proven-styles-snoopy.md
⚠️ 关键经验： 不要在prompt中过度约束视觉细节（颜色比例、构图位置、角色姿势），否则会严重降低多样性。只描述情绪和内容，让AI自由发挥

2. Manga Educational — 学習漫画风

Philosophy: Japanese educational manga (学習漫画) — a character GUIDES you through the concept with reactions and drama
Colors: Bright and warm palette, white bg with selective color panels, screen-tone gray for emphasis areas
Ratio: 60% illustration / 30% text (in bubbles) / 10% effects
Typography: Bold manga-style titles with impact, body text in speech/thought bubbles, onomatopoeia as decorative elements. Size contrast 3:1
Composition: Dynamic manga panel layouts (3-5 panels per slide), character reactions drive emphasis, speed lines for energy, dramatic angles
Visual language: Expressive anime-style characters, reaction faces (surprise, confusion, eureka!), manga effects (sweat drops, sparkles, speed lines), panel borders with varied thickness
Reference: "Like a 'Manga Guide to Statistics' page — a character walks you through the concept, reacting with surprise and delight"

3. Ligne Claire Comics — 清线漫画风

Philosophy: Hergé's Tintin tradition — maximum information clarity through visual restraint
Colors: White/cream (#FFFDF7) bg, black (#000000) outlines, flat saturated fills (3-5 solid colors, no gradients)
Ratio: 70% clean bg / 20% illustration / 10% text
Typography: Hand-lettered feel for titles, clean sans-serif for body. Speech bubbles for key quotes. Title:body = 2.5:1
Composition: Panel-based layouts (2-4 panels per slide), sequential left-to-right reading flow, clear gutters between panels
Visual language: Uniform-weight outlines, flat colors without shading or hatching, no gradients, precise details but zero visual noise
Reference: "Like a Tintin page explaining a concept — every panel advances understanding, nothing is decorative"

4. Neo-Pop Magazine — 新波普杂志风

Philosophy: Youth media / streetwear brand aesthetic, bold and playful
Colors: Cream (#FFF8E7) bg, black (#000000) text, color-blocking with hot pink (#FF1493) + cyan (#00CED1) + yellow (#FFD700)
Ratio: 50% bg / 25% color blocks / 25% content
Typography: Headlines 40-50% of slide area (typography AS the visual), thick black borders around text blocks, 10:1 size ratio vs body
Composition: Modular color blocks with "controlled chaos", stacked asymmetric layouts, thick borders
Visual language: Pixel-art 8-bit icons, cutout photography, speech bubbles, bold graphic surfaces
Reference: "Like a Supreme lookbook meets a HYPEBEAST article — treats typography as graphic art"

第二梯队（推荐，特定场景效果好）：

5. Whiteboard Sketch — xkcd白板手绘风

Philosophy: xkcd meets a professor's whiteboard — extreme minimalism forces focus on the idea itself
Colors: White (#FFFFFF) bg, black (#000000) ink, ONE accent color for emphasis (red #FF4444 or blue #4488FF)
Ratio: 85% white space / 10% sketch / 5% accent highlight
Typography: Hand-drawn/handwritten feel for everything, rough uneven baselines, arrows and annotations everywhere. Key numbers can be large (60pt+)
Composition: Freeform whiteboard layout, hand-drawn arrows connecting concepts, diagrams and stick figures, informal and alive
Visual language: Stick figures, hand-drawn charts and graphs, wobbly lines, annotation arrows, circled keywords, equation-style layouts
Reference: "Like an xkcd 'What If?' explanation — simple drawings that make complex ideas instantly click"

6. Soviet Constructivism — 苏联构成主义

Philosophy: Revolutionary propaganda poster — power through geometry and limited color
Colors: Revolutionary red (#CC0000) 40% + black (#1A1A1A) 25% + cream white (#F5E6D3) 30%
Typography: All text rotated 15-30 degrees, NO horizontal lines, bold condensed
Composition: Diagonal wedge from bottom-left to top-right, geometric shapes growing small to large (visual crescendo)
Visual language: NO gradients, pure flat fills + sharp edges, three-color limit, propaganda poster energy
Reference: "Like a 1920s Rodchenko poster — power, urgency, and geometric precision"

7. Warm Narrative — 温暖叙事风

Philosophy: Friendly storytelling, like a TED talk visual or Airbnb pitch deck
Colors: Warm cream (#FDF6EC) bg, dark charcoal (#3D3D3D) text, coral (#E17055) accent
Ratio: 60% warm bg / 25% content / 15% illustration
Typography: Headlines bold and warm, 3:1 ratio to body. Short sentences, not bullets
Composition: Illustration occupies 40-50% of slide, text wraps around visuals, rounded shapes
Visual language: Flat vector illustrations with warm palette, people-centric imagery, storytelling flow
Reference: "Like a Mailchimp or Notion brand presentation — approachable and human"

更多风格（第二/三梯队） 详见 references/proven-styles-gallery.md，包括：The Oatmeal信息图漫画、敦煌壁画、浮世绘、孔版印刷Risograph、等轴测Isometric、Bauhaus包豪斯、工程蓝图Blueprint、复古广告Vintage Ad、达达拼贴Collage、像素画Pixel Art

🎨 Custom Character Style (User-Defined)

Users may want to reference specific cartoon/anime aesthetics. When a user says "do it in Doraemon style" or "like Studio Ghibli", treat this as a style reference, not a request to draw copyrighted characters. Build a custom Design System by extracting the visual DNA of that style.

How to convert a character reference into a Design System:

User says	Extract these visual traits
"Doraemon style"	Round shapes, bright primary blue + white + red, simple backgrounds, cute proportions, magical gadget reveals
"Studio Ghibli"	Watercolor textures, natural greens and sky blues, detailed backgrounds with simple characters, warmth and wonder
"Calvin and Hobbes"	Dynamic ink brushwork, expressive motion lines, philosophical contrast between fantasy and reality, lush outdoor scenes
"One Piece manga"	Bold dynamic lines, exaggerated proportions, dramatic action poses, high energy, thick outlines
"Crayon Shin-chan"	Crude crayon-like lines, flat bright colors, comedic proportions, everyday scenarios made absurd
"Adventure Time"	Geometric simple shapes, pastel candy colors, thin outlines, whimsical surreal backgrounds

Template for custom style:

[User Style]: "[reference name]"
→ Shape language: [round/angular/geometric/organic]
→ Line quality: [thin uniform / thick varied / sketchy / brushwork]
→ Color palette: [specific colors extracted from that aesthetic]
→ Character style: [proportions, expressiveness level]
→ Background treatment: [detailed/minimal/abstract]
→ Emotional tone: [warm/energetic/philosophical/surreal]

Typography Rules (All Presets)

Max 2 font families (1 heading + 1 body)
Heading: bold, personality — ≥36pt (trend: even larger, as graphic surface)
Body: clean, readable — ≥18pt
Chinese: system default (PingFang SC / Microsoft YaHei)
Key principle: Typography is a DESIGN ELEMENT, not just an information container

✅ Checkpoint 2 (Guided + Collaborative)

Ask the user to pick one of the 3 proposed design systems, or describe their own preference. Show the full description including philosophy, visual language, and reference.

Step 3: Build Slides

Step 3-A: HTML + Selective Illustrations (Path A)

Generate AI illustrations for key slides, then create HTML slide files.

Which slides need illustrations? Prioritize:

Cover slide — always. Sets the visual tone.
Key insight slides — the "aha moment" slides benefit most.
Closing slide — optional but impactful.
Data-heavy slides — charts/diagrams instead of AI art.

Illustration Generation — use nano-banana-pro skill:

export $(grep GEMINI_API_KEY ~/.claude/.env) && \
uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
  --prompt "[description]" \
  --filename "[timestamp]-slide-[N]-[name].png" \
  --resolution 2K

Base Style Prompt — define ONE style suffix, append to every illustration:

[Base Style]: flat vector illustration, [palette background color] background,
[accent color] highlight elements, clean minimalist aesthetic,
professional presentation style, no text in image

Per-slide prompt = [specific content] + [Base Style]

Key rules:

Always include "no text in image" — text will be added as editable elements
Use descriptive paragraphs, not keyword lists
Specify hex colors explicitly
Use "flat vector" / "flat illustration" for consistency

Embedding in HTML slides:

<!-- Side illustration (recommended) -->
<div class="left"><!-- text content --></div>
<div class="right"><img src="illustration.png" style="width: 280pt; height: 280pt;"></div>

<!-- Background illustration -->
<body style="background-image: url('illustration.png'); background-size: cover;">

✅ Checkpoint 3-A (Guided: preview 2-3 key illustrations; Collaborative: every one)

Show generated illustrations. Ask: Approve / regenerate / style consistent?

Step 3-B: Full AI Slide Generation (Path B)

Generate EVERY slide as a complete AI image — layout, text, visuals, all in one.

⚠️ THE #1 MISTAKE: Over-constraining the prompt with layout details and visual restrictions. More constraints = LESS creativity and diversity. The AI generates best when given mood + reference + content, NOT specific positions, color ratios, or character restrictions.

The Golden Rule of AI Image Prompts

SHORT prompts > LONG prompts. A 3-sentence prompt describing mood and content produces better results than a 30-line prompt specifying every visual detail. Specifically:

DON'T (kills diversity)	DO (enables creativity)
Specify color ratios (60%/25%/15%)	Describe the mood ("warm like a Sunday comic page")
Dictate layout positions ("title centered, image on right")	Reference a specific aesthetic ("Peanuts comic strip")
Restrict characters ("NOT Snoopy — an original character")	Let AI interpret the style naturally
List every visual element to include	Describe what the viewer should FEEL
Repeat the base style in every per-slide prompt	Define base style once, keep per-slide prompts short

Base Style Prompt — Keep it SHORT

Define a base style once, append to every slide. Keep it under 5 lines. The base style sets the mood; per-slide prompts add the content.

[Base Style]:
VISUAL REFERENCE: [Specific art/design aesthetic in one sentence]
CANVAS: 16:9 aspect ratio, 2048x1152 pixels, high quality rendering.
COLOR SYSTEM: [Describe the mood/feel of colors, not exact ratios]

Example (good — concise):

VISUAL REFERENCE: Charles Schulz Peanuts comic strip — warm, philosophical, charming.
Characters include round-headed kids, a lovable beagle dog, and a small yellow bird.
CANVAS: 16:9 aspect ratio, 2048x1152 pixels, high quality rendering.
COLOR SYSTEM: Warm cream/newspaper tone background, soft muted pastels, warm ink lines.

Anti-pattern (bad — over-specified): Do NOT include typography sizes, color ratios, composition percentages, margin specifications, or visual weight distributions in the base style. These constraints reduce diversity without improving quality.

Per-Slide Prompt Structure

Keep per-slide prompts short and focused. Do NOT repeat base style details or over-specify visual layout.

Create a [style] slide about [topic].

[Base Style]

DESIGN INTENT: [1 sentence — what the viewer should FEEL]

TEXT TO RENDER:
- Title: "[exact text]"
- Body: "[exact text]"

[Optional: 1-2 sentences describing mood or scene. Let AI decide composition.]

Example — GOOD vs BAD

BAD (traditional PPT — boring):

Design a professional presentation slide.
Professional presentation slide, 16:9 aspect ratio, 2048x1152 pixels.
Dark navy background, light gray text, gold accent.
Slide type: content. Layout: Title at top-left, two columns below.
Title: "看涨期权收益结构"
Body: "行权价: 100元, 权利金: 10元"
Visual: a line chart showing call option payoff

→ Result: Generic PPT that could come from any template

GOOD (magazine-level — stunning):

Create a slide that feels like a Bloomberg terminal data visualization
brought to life as editorial art.

VISUAL REFERENCE: Bloomberg Businessweek data feature meets cinematic lighting.
CANVAS: 16:9, 2048x1152, sharp rendering.
COLOR SYSTEM: Deep black (#0A0A0A) background 75%, white text 15%,
gold (#BF9A4A) accent 10%. The gold represents profit — it should GLOW.
TYPOGRAPHY: The number "110" rendered at 100pt as the dominant visual anchor
(the break-even point IS the story). Supporting text at 14pt, muted gray.

DESIGN INTENT: The viewer should instantly FEEL the asymmetry of options —
limited downside, unlimited upside. The visual must make this visceral,
not just informational.

TEXT TO RENDER:
- Hero metric: "110" (giant, gold, the break-even price)
- Title: "盈亏平衡点" (medium, white, above the number)
- Left data: "行权价 100" "权利金 10" (small, gray, understated)
- Insight: "亏损有底 盈利无限" (accent color, bottom)

VISUAL NARRATIVE: A single golden curve emerges from the left side of the slide,
flat and muted in gray at -10 (the maximum loss), then suddenly bending upward
at the strike price, transitioning from gray to brilliant gold as it rises
into the profit zone. The curve should feel like a ray of light breaking
through darkness. The profitable area above zero glows with warm gold
atmospheric lighting, like sunrise. The chart has NO grid lines, NO axes labels
cluttering the visual — just the pure, dramatic curve and the giant "110"
floating at the inflection point.

→ Result: An editorial data visualization that tells a story

Key Rules for Path B Prompts

Prompt Quality Checklist (verify before every generation):

Visual Reference — Does the prompt name a specific art style or publication? (NOT just "professional" or "modern")
Mood, not Layout — Does the prompt describe what the viewer should FEEL, not where elements should be PLACED?
Text Content — Are all texts to render listed clearly and accurately?
Short Enough — Is the prompt concise? Long prompts with detailed specs REDUCE diversity. Remove anything the AI can decide on its own.
NO Micro-Management — No hex color ratios, no typography sizes, no composition percentages, no character pose instructions.

Technical Rules:

Always specify resolution: 2048x1152 (2K, 16:9) for crisp text
Include ALL text verbatim — AI must render exact words
中文优先: slide上的文字一律用中文，仅保留必要英文术语
Chinese text tip: Keep titles short (≤8 characters) for best rendering
Use descriptive paragraphs, not keyword lists
Generate in parallel: Run 3-5 slide generations concurrently for speed
Consistency: The Base Style is applied to EVERY slide. It's a system, not a suggestion

Generation command (same tool, but full-slide prompts):

export $(grep GEMINI_API_KEY ~/.claude/.env) && \
uv run ~/.claude/skills/nano-banana-pro/scripts/generate_image.py \
  --prompt "[full slide prompt]" \
  --filename "slide-[NN]-[name].png" \
  --resolution 2K

Quality check after generation:

Text accuracy — verify all Chinese/English text rendered correctly
Layout — elements positioned as described
Style consistency — colors and design language match across slides
If a slide has text errors → regenerate with adjusted prompt (simplify text or shorten)

✅ Checkpoint 3-B (Guided: preview all slides as a set; Collaborative: approve each)

Show ALL generated slide images to the user. Ask:

Text readable and accurate?
Visual style consistent across slides?
Any slides to regenerate?

Step 4: PPTX Assembly

4-A: html2pptx Workflow (Path A)

Create HTML files per slide, convert with html2pptx.js:

const pptxgen = require('pptxgenjs');
const html2pptx = require(process.env.HOME + '/.agents/skills/pptx/scripts/html2pptx.js');

const pptx = new pptxgen();
pptx.layout = 'LAYOUT_16x9';
await html2pptx('slide1.html', pptx);
await html2pptx('slide2.html', pptx);
await pptx.writeFile({ fileName: 'output.pptx' });

HTML rules (from pptx skill):

Body dimensions: width: 720pt; height: 405pt (16:9)
ALL text must be in <p>, <h1>-<h6>, <ul>, <ol> tags
Backgrounds/borders only on <div> elements
No CSS gradients — pre-render as PNG with Sharp
Use web-safe fonts only (Arial, Helvetica, Georgia, etc.)
Images: <img src="illustration.png" style="width: Xpt; height: Ypt;">

Known issue: Chinese characters in file paths can break image loading. Use symlinks to ASCII paths if needed:

ln -sf "/path/with/中文/" /tmp/ascii-path

4-B: Image Assembly (Path B)

Assemble generated slide images into PPTX using create_slides.py:

uv run ~/.claude/skills/image-to-slides/scripts/create_slides.py \
  slide-01-cover.png slide-02-intro.png slide-03-definition.png ... \
  --layout fullscreen \
  --bg-color 000000 \
  -o output.pptx

Recommended layout for Path B: fullscreen — images fill the entire slide since they already contain all layout, text, and visuals.

Layout	Use case
`fullscreen`	AI-generated full-page slides (Path B default)
`title_above`	Image + editable title (hybrid approach)
`title_left`	Split: text + visual
`center`	Centered image with padding
`grid`	Multiple images per slide

Step 5: Preview & Polish

Preview

Path A: Screenshot 3-4 key HTML slides with Playwright:

npx playwright screenshot "file:///path/to/slide.html" preview.png \
  --viewport-size=960,540 --wait-for-timeout=1000

Path B: Show the generated slide images directly (they ARE the slides). Use Read tool to display 3-4 key PNGs.

✅ Checkpoint 4 (All modes)

Show preview to the user. The PPTX file is ready — ask:

Any slides to adjust?
Ready to open in Keynote/PowerPoint?

Final Polish (in Keynote/PowerPoint)

Transitions and animations
Speaker notes
Brand logo placement
Path A: Final text adjustments (editable)
Path B: Text NOT editable — if text errors found, regenerate the slide image

Design Quick Reference

5/5/5 rule: ≤5 words/line, ≤5 bullets/slide, ≤5 text-heavy slides in a row

Cognitive load: One idea per slide. ~1 min per slide. Slides complement speech, never duplicate it.

Visual hierarchy: F/Z-pattern reading flow. Title:body size ≈ 3:1. Every slide should have a visual element.

Detailed references:

references/proven-styles-gallery.md — 17 tested visual styles with tiered recommendations
references/proven-styles-snoopy.md — Snoopy/Peanuts style detailed per-slide templates
references/prompt-templates.md — Content generation and image prompts
references/design-principles.md — Full design framework, color palettes, typography

Related Skills

Skill	Role
`pptx`	Advanced PPTX creation/editing (html2pptx, templates)
`nano-banana-pro`	AI illustration generation (Gemini 3 Pro Image)
`multi-model`	External AI for content drafting

Output

.pptx files compatible with PowerPoint, Keynote, Google Slides
Web-safe fonts for cross-platform compatibility
AI illustrations as separate PNG files (reusable)

ai-presentation