Morph Ppt 3d

Model Discovery Flow (when user has no model)

When the user gives a topic but no .glb file, proactively help them find a matching model instead of just listing websites.

Step 1: Understand the topic and suggest model direction

Based on the user's topic, suggest what kind of 3D model would work:

Tell the user: "你的主题是 [X]，建议用 [具体模型描述] 的 3D 模型。我推荐几个免费下载来源："

Step 2: Search for models (agent-driven)

Proactively search for models on behalf of the user. Don't just list websites — actually find candidates.

Search strategy (try in order):

1. Web search for free GLB models matching the topic:

   Search: "[topic keyword] 3d model glb free download"
   Example: "shiba dog 3d model glb free download"
   

2. Sketchfab API (no auth needed for search):

   curl -s "https://api.sketchfab.com/v3/search?type=models&q=[keyword]&downloadable=true&archives_flavours=glb" \
     | python3 -c "
   import json, sys
   data = json.load(sys.stdin)
   for m in data.get('results', [])[:5]:
       print(f\"Name: {m['name']}\")
       print(f\"URL: https://sketchfab.com/3d-models/{m['slug']}-{m['uid']}\")
       print(f\"Likes: {m.get('likeCount', 0)}, License: {m.get('license', {}).get('label', 'unknown')}\")
       print()
   "
   

3. Poly Pizza (direct GLB download, all free):

   # Search results page — parse for download links
   curl -s "https://poly.pizza/api/search/[keyword]" 2>/dev/null
   

4. Khronos glTF-Sample-Assets (guaranteed to work, always available):

   # Direct download — no auth, no API, always works
   curl -L -o model.glb "https://raw.githubusercontent.com/KhronosGroup/glTF-Sample-Assets/main/Models/[ModelName]/glTF-Binary/[ModelName].glb"
   

   Available models: Duck, Fox, Avocado, BrainStem, CesiumMan, DamagedHelmet, FlightHelmet, Lantern, Suzanne, WaterBottle, etc.

Step 3: Present candidates to user for confirmation

Show the user 2-3 model options with:

• Model name and source
• Preview link (Sketchfab URL or description)
• License info
• Why this model fits their topic

Example response:

根据你的主题"柴犬品牌"，我找到了这些模型：

1. 🐕 Shiba Inu (Sketchfab)
   链接：https://sketchfab.com/3d-models/shiba-xxx
   授权：CC BY 4.0（免费可商用）
   推荐理由：高质量柴犬模型，表情可爱

2. 🐶 Low Poly Dog (Poly Pizza)
   链接：https://poly.pizza/m/xxx
   授权：CC0（完全免费）
   推荐理由：低多边形风格，适合简洁设计

3. 🦊 Fox (Khronos 官方样例)
   可直接下载，保证兼容
   推荐理由：狐狸和柴犬外形相似，作为备选

你选哪个？确认后我直接下载开始做。

Wait for user confirmation before downloading. Do not download without asking.

Step 4: Download the confirmed model

After user confirms, download directly:

# For Sketchfab (if user has the download URL)
curl -L -o model.glb "[download_url]"

# For Khronos samples (always works)
curl -L -o model.glb "https://raw.githubusercontent.com/KhronosGroup/glTF-Sample-Assets/main/Models/Fox/glTF-Binary/Fox.glb"

After download, verify:

• File exists and is not empty
• File extension is .glb
• File size is under 50MB

If Sketchfab requires login to download, tell the user:

"这个模型需要在 Sketchfab 登录后下载。你可以去页面下载 .glb 文件，然后上传给我。或者我用 Khronos 官方样例先做一版演示？"

Step 5: When user says "随便" / "你定" / "先做个演示"

Don't just grab a random model. First guide the user to clarify their PPT topic:

好的！模型我来搞定，但先确认一下你的 PPT 主题方向，这样我找的模型才能配合内容：

1. 🎮 科技/产品 — 耳机、手机、机器人...
2. 🐾 动物/角色 — 可爱宠物、卡通人物...
3. 🏗️ 建筑/空间 — 房屋、室内、城市...
4. 🍕 食物/生活 — 美食、日用品...
5. 🚀 其他 — 告诉我你的想法

选一个方向，或者直接说个主题词也行。

After user confirms a direction, THEN search and recommend models.

Only if user explicitly says "真的随便" / "什么都行" / insists on no preference, use a built-in model:

Built-in models (bundled with the skill, no download needed):

# Find and copy built-in model to working directory (cross-platform)
MODEL_FOUND=""
for BASE in "$HOME/Library/Application Support" "$APPDATA" "$XDG_CONFIG_HOME" "$HOME/.config"; do
  [ -z "$BASE" ] && continue
  F="$(find "$BASE" -path "*/builtin-skills/morph-ppt-3d/models/shiba.glb" -print -quit 2>/dev/null)"
  if [ -n "$F" ]; then MODEL_FOUND="$F"; break; fi
done
if [ -n "$MODEL_FOUND" ]; then
  cp "$MODEL_FOUND" ./model.glb
else
  # Fallback: download from Khronos
  curl -L -o model.glb "https://raw.githubusercontent.com/KhronosGroup/glTF-Sample-Assets/main/Models/Duck/glTF-Binary/Duck.glb"
fi

If the built-in model doesn't fit the user's topic, fall back to Khronos samples:

curl -L -o model.glb "https://raw.githubusercontent.com/KhronosGroup/glTF-Sample-Assets/main/Models/Duck/glTF-Binary/Duck.glb"

Step 6: When user wants to find models themselves

Give specific website links with step-by-step guidance:

推荐的 3D 模型网站：

1. Sketchfab (最大的 3D 模型平台)
   • 链接：https://sketchfab.com/search?q=[关键词]&type=models&downloadable=true
   • 筛选步骤：搜索关键词 → 勾选 "Downloadable" → 格式选 "glTF" → 按 "Likes" 排序
   • 下载时选 glTF (.glb) 格式
   • 注意：部分模型需要免费注册后才能下载
2. Poly Pizza (全免费低多边形)
   • 链接：https://poly.pizza/
   • 特点：全部免费 CC0 授权，直接点 Download 就是 .glb
   • 适合：简约风格、卡通风格的 PPT
3. Sketchfab 热门分类直达
   • 动物：https://sketchfab.com/search?q=animal&type=models&downloadable=true
   • 食物：https://sketchfab.com/search?q=food&type=models&downloadable=true
   • 科技：https://sketchfab.com/search?q=gadget&type=models&downloadable=true
   • 建筑：https://sketchfab.com/search?q=architecture&type=models&downloadable=true
4. Free3D (综合免费模型站)
   • 链接：https://free3d.com/3d-models/glb
   • 注意：需确认授权类型
5. TurboSquid Free (专业模型站免费区)
   • 链接：https://www.turbosquid.com/Search/3D-Models/free/glb

下载后把 .glb 文件发给我就行。如果下载的是 .gltf（文件夹），需要用 Blender 转成 .glb。

Step 7: When user gives keywords and asks agent to search

Remind about token cost before searching:

我可以帮你搜索，不过在线搜索会消耗一些额外的对话额度 (token)。你想：

A. 我来搜 — 我用 Sketchfab API 搜索并推荐 2-3 个（消耗少量 token） B. 你自己找 — 我给你搜索链接和筛选教程，你挑好发给我（不消耗额外 token）

选 A 还是 B？

If user chooses A, proceed with Step 2 (agent-driven search). If user chooses B, proceed with Step 6 (self-service guidance).

License reminder

Always remind before confirming download: "下载前请确认模型授权。CC0 / CC BY 可免费使用；CC BY-NC 仅限非商用。"

Visual Design System (4.0 enrichment)

morph-ppt provides the base design rules. This section adds concrete palettes, font pairings, and layout quality rules from PPT Creator to give the AI more variety and stronger guardrails.

Color Palettes (pick one per deck, or blend)

Choose a palette that matches the topic mood — don't default to generic blue.

Rules:

• One color dominates (60-70% visual weight), 1-2 supporting tones, one accent
• On light backgrounds: use Body Text color for copy, Muted for captions
• On dark backgrounds: use Secondary or FFFFFF for copy, Muted for captions
• You can also use morph-ppt's reference/styles/ for additional inspiration — these palettes are complementary, not replacements

Font Pairings (pick one per deck)

Hard Rules (mandatory, no exceptions)

H4 — Body text minimum 16pt: All body text, card content, and bullet points must be >= 16pt. "Content doesn't fit" is not an excuse — reduce text, split slides, or reduce card count instead. Exceptions: chart axis labels (<=12pt), short sublabels (<=14pt, max 5 words), footnotes.

H6 — Dark background contrast: When slide background brightness < 30% (e.g. 1E2761, 36454F, 000000), ALL body text, card content, chart labels, and icon fills MUST use white (FFFFFF) or near-white (brightness > 80%). Never use mid-gray or muted colors as body text on dark backgrounds.

H7 — Speaker notes required: Every content slide (not title/closing) MUST have speaker notes. Use:

officecli add deck.pptx '/slide[N]' --type notes --prop text="..."

Visual Element Checkpoint

Every 3 content slides, at least 1 must contain a non-text visual element:

Text-only slides are only allowed for: quotes, code examples, pure tables.

3D Model Insertion Rules

Add model fresh on every slide — NEVER clone

morph_clone_slide copies the model as frozen XML. The cloned model cannot Morph. Each slide must call add --type 3dmodel independently with the same name prop.

⚠️ CRITICAL: If you clone a slide that already has a 3D model, the old model XML is copied too. This creates TWO model3d elements with the same name on the new slide. PowerPoint cannot handle this conflict and will delete the model content during repair.

If you must clone a slide for scene actors, immediately remove the cloned model before adding a new one:

# After cloning slide 1 to slide 2:
officecli remove deck.pptx '/slide[2]/model3d[1]'  # remove the frozen clone
officecli add deck.pptx '/slide[2]' --type 3dmodel ...  # add fresh model

Recommended approach: Do NOT clone slides with 3D models at all. Create all slides empty first, then add models fresh on each.

# Slide 1
officecli add deck.pptx '/slide[1]' --type 3dmodel \
  --prop path=model.glb --prop 'name=!!model-hero' \
  --prop x=16cm --prop y=1cm --prop width=16cm --prop height=16cm \
  --prop roty=0

# Slide 2
officecli add deck.pptx '/slide[2]' --type 3dmodel \
  --prop path=model.glb --prop 'name=!!model-hero' \
  --prop x=0.5cm --prop y=1cm --prop width=18cm --prop height=17cm \
  --prop roty=50

Controllable properties

Do NOT manually set

• meterPerModelUnit — auto-computed from GLB bounding box
• preTrans — auto-computed for model centering
• camera depth/position — auto-computed to fit the model
• Never use raw-set on any 3D transform parameter

Model-Content Layout

Core Principle: Model IS the Subject

The model must feel like the protagonist of the presentation, not a sidebar decoration. Text supports the model; the model does not decorate the text.

Size Contrast Rule (MANDATORY)

Adjacent slides must have a model area ratio >= 1.5x or <= 0.67x. Compute area as width × height. If slide N model is 16×15=240 cm², slide N+1 must be >= 360 or <= 160.

Never use similar sizes on consecutive slides. This is the single most important rule for visual energy.

Layout Patterns (6 types)

A — Model right, content left (content pages) Content at x=1-14cm. Model at x=15-20cm, width 14-18cm.

B — Model left, content right (alternate with A) Model at x=0-2cm, width 14-18cm. Content at x=18-32cm.

C — Model centered, text overlay (title/closing) Model centered large (18-24cm). Text at slide top or bottom.

D — Model small corner, content dominant (data pages) Model 5-10cm in any corner. Content fills the rest.

E — Model as backdrop (impact/quote pages) Model XL (28-36cm), centered, partially cropped by slide edges. Text overlaid directly on top of model area with high-contrast color. The model becomes the "canvas" — text lives inside the model's space.

# Pattern E: model fills slide as backdrop
officecli add deck.pptx '/slide[N]' --type 3dmodel \
  --prop path=model.glb --prop 'name=!!model-hero' \
  --prop x=-2cm --prop y=-2cm --prop width=38cm --prop height=24cm \
  --prop roty=45 --prop rotx=10

# Text overlaid on model
officecli add deck.pptx '/slide[N]' --type shape \
  --prop 'name=#sN-quote' --prop text="Key insight here" \
  --prop x=3cm --prop y=7cm --prop width=28cm --prop height=5cm \
  --prop size=44 --prop bold=true --prop color=FFFFFF --prop fill=none

F — Model bleed edge (transition/teaser pages) Model partially off-screen (negative x or y, or x+width > 33.87cm). Only part of the model visible — implies more beyond the frame.

# Pattern F: model bleeds off right edge
officecli add deck.pptx '/slide[N]' --type 3dmodel \
  --prop path=model.glb --prop 'name=!!model-hero' \
  --prop x=20cm --prop y=-1cm --prop width=24cm --prop height=22cm \
  --prop roty=70

Layout Progression

Never repeat the same pattern on consecutive slides. Example:

Slide 1: C (centered hero, L)
Slide 2: E (backdrop close-up, XL)   ← 1.5x+ area jump
Slide 3: A (model right, M)          ← pull back
Slide 4: F (bleed edge, L)           ← push in
Slide 5: D (small corner, S)         ← dramatic pull back
Slide 6: B (model left, M)           ← grow
Slide 7: C (centered closing, L)     ← push in

Text Layout Safety (MANDATORY)

Text boxes must never overlap each other or the model frame.

Rules:

1. Title and body must not collide. If a title wraps to 2 lines, the body y must account for the title's actual height, not the planned height. Safe formula: body_y = title_y + title_height + 0.5cm
2. Fixed-height text boxes are dangerous. If text content is longer than expected, it will overflow invisibly. Use generous heights: title 3-4cm, body 6-8cm, bullets 8-10cm.
3. Model frame and text boxes: gap >= 1cm. Calculate: if model is at x=15cm, text x + width must be <= 14cm.
4. On Pattern C (centered model + text overlay): text goes at slide top (y=0.5-2cm) or bottom (y=14-17cm), NOT in the vertical middle where the model lives (y=3-13cm).
5. After building each slide, verify coordinates:

   officecli get deck.pptx '/slide[N]' --depth 1
   # Check: no two shapes share overlapping x/y/width/height ranges
   

Model Bleed Guidelines

Not every model looks good when cropped. Bleed (Pattern E/F) works best for:

• ✅ Symmetric objects (spheres, helmets, bottles) — any crop looks intentional
• ✅ Large flat surfaces (cars, buildings) — partial view implies scale
• ✅ When cropping non-critical parts (background, base, stand)

Bleed does NOT work for:

• ❌ Character/animal models — cropping ears, tails, or limbs looks broken
• ❌ Small detailed models — cropping loses the detail you want to show
• ❌ When the cropped part is the most recognizable feature

For character/animal models (like shiba, fox, duck): keep the full model visible on all slides. Use size changes (L→M→S) for rhythm instead of bleed cropping. Use rotx for angle variety instead.

Camera Language

Three tools work together: roty (orbit), rotx (tilt), width/height (zoom).

Shot Types (use >= 3 different per deck)

Content-Driven Camera

Match the shot to what the slide talks about:

• "Front design" → Close-up, roty=0, XL cropped
• "Side profile" → Side, roty=90, M
• "Internal structure" → Bird's eye, roty=30, rotx=35, M
• "Power/authority" → Low angle, roty=20, rotx=-20, L
• "Data & specs" → Over-the-shoulder, roty=60, S in corner

Rotation Rules

1. Adjacent roty delta: 30-90° (< 30 = jitter, > 90 = disorienting)
2. Overall roty direction must be consistent (no back-and-forth)
3. rotx range: -25 to +40. Adjacent rotx delta <= 20
4. Total arc across deck: 180-360° (show the model from all sides)

Example Shot Plan

Workflow Integration with morph-ppt

Phase 2 additions (Planning)

In brief.md, add a Model Choreography Table:

Verify the area ratio rule (>= 1.5x between adjacent rows) before proceeding to build.

Phase 3 additions (Build)

Since models cannot be cloned, the build script differs from standard morph-ppt:

1. Create all slides first (with background + morph transition)
2. Add scene actors (!!scene-*) on slide 1, then clone slides for morph continuity
3. Add 3D model fresh on EACH slide (same name, different roty/position)
4. Add content shapes per slide, ghost previous content

model_positions = [
    {"slide": 1, "x": "7cm",  "y": "0.5cm", "w": "20cm", "h": "17cm", "roty": 30},
    {"slide": 2, "x": "-2cm", "y": "-2cm",  "w": "38cm", "h": "24cm", "roty": 0},
    {"slide": 3, "x": "16cm", "y": "1cm",   "w": "15cm", "h": "14cm", "roty": 80},
    # ...
]
for pos in model_positions:
    run("officecli", "add", OUTPUT, f"/slide[{pos['slide']}]", "--type", "3dmodel",
        "--prop", f"path={MODEL}", "--prop", "name=!!model-hero",
        "--prop", f"x={pos['x']}", "--prop", f"y={pos['y']}",
        "--prop", f"width={pos['w']}", "--prop", f"height={pos['h']}",
        "--prop", f"roty={pos['roty']}")

Phase 4 additions (Verification)

After standard morph verification, additionally check:

• Each slide has exactly one model3d element
• All models share the same name prop
• Adjacent slides have model area ratio >= 1.5x or <= 0.67x
• No two consecutive slides use the same layout pattern

File Placement Rule

All files must be in the same working directory.

Deliverables (exactly 4 files, no more):

• .glb model file (the 3D model used in the deck)
• Output .pptx
• Build script (re-runnable)
• brief.md

Do NOT create additional files such as outline.md, quality-report.md, test-report.md, etc. All planning goes in brief.md, all verification output goes to stdout. Extra files confuse users.

Do not scatter model files across unrelated paths.