Morph Ppt 3d

Morph PPT — 3D Extension

This skill extends morph-ppt. All morph-ppt rules (naming, ghosting, design, verification) apply in full. This file covers 3D-specific additions and an enriched design system combining morph-ppt aesthetics with concrete color palettes, font pairings, and layout quality guardrails.

运行环境与交付链（本系统专用，先读这一段）

officecli 已经预装在沙盒镜像的 /usr/local/bin/officecli，不要联网安装、不要 curl。 bash 工具直接调用即可，自检用 officecli --version。

所有命令都在沙盒里执行，文件一律落在 /workspace/ 下。一次完整作业 = 3 步 + 1 步交付：

# Step 1（仅当用户上传过原始文件）把 artifact 送进沙盒
sandbox_put_artifact(artifact_id="<用户文件的 file_id>", dest_path="/workspace/in.docx")

# Step 2 用 bash 跑 officecli —— 本文件下面讲的全部命令都在这一步里
bash("officecli create /workspace/out.docx")

# Step 3 把成稿从沙盒取出、登记成 artifact，拿到新的 file_id
sandbox_get_artifact(src_path="/workspace/out.docx", name="终稿.docx")

# Step 4（必做）交付给用户
pin_to_workspace(file_ids=["<新 file_id>"])

铁律：

• sandbox_put_artifact = 放进沙盒，sandbox_get_artifact = 从沙盒取出，别搞反；两个路径参数都必须以 /workspace/ 开头。
• 拿到 file_id 却不 pin_to_workspace 就等于没交付，用户看不到文件。多个产物一次性放进同一个列表，别分多次调用。
• 下文示例里的 $FILE 和相对路径，都要换成 /workspace/ 下的绝对路径。
• 下文的可视化质检（view <file> screenshot / view <file> html）依赖沙盒内的 Chromium，镜像已预装；截图失败就按下文 fallback 走 view html，并如实标注「未做视觉验证」，不要谎报 PASS。
• 导出 PDF 不要用 officecli view <file> pdf —— 那个模式依赖本系统没有安装的 exporter 插件，必然失败。改用沙盒里的 LibreOffice： soffice --headless --convert-to pdf --outdir /workspace /workspace/out.docx。.xlsx / .pptx 换掉扩展名同理，但要沙盒里装了 LibreOffice Calc / Impress 才转得动；转换报错就如实告诉用户 PDF 没导出成功、主产物仍是原文件，不要谎报已导出。

Use when

• User wants a .pptx with a .glb 3D model and Morph transitions.

3D Model Compatibility Gate (before generation)

1. Only .glb is supported. If user provides .fbx / .obj / .blend / .usdz / .gltf, ask them to convert to .glb first (e.g. via Blender export).
2. If user has no model, follow the Model Discovery Flow below.
3. All files (.glb, .pptx, build script) must be in the same working directory.

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: "Your topic is [X]. I suggest using a 3D model of [description]. Here are some free sources to find one:"

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: "fox 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:

Based on your topic "fox mascot", here are some models I found:

1. Fox (Khronos sample)
   Direct download, guaranteed compatible
   Why: clean fox model, good for mascot/character decks

2. Low Poly Fox (Poly Pizza)
   URL: https://poly.pizza/m/xxx
   License: CC0 (completely free)
   Why: low-poly style, good fit for clean minimal design

3. Cartoon Fox (Sketchfab)
   URL: https://sketchfab.com/3d-models/fox-xxx
   License: CC BY 4.0 (free, commercial use ok)
   Why: expressive face, high detail

Which one do you want? I'll download it and start building.

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:

"This model requires a Sketchfab login to download. You can grab the .glb file from the page and share it with me. Or I can use a Khronos sample model for a demo version first?"

Step 5: When user says "anything" / "you decide" / "just make a demo"

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

Sure! I'll handle the model — but let me confirm the topic direction first so the model matches the content:

1. Tech/Product — headphones, phone, robot...
2. Animal/Character — cute pet, cartoon character...
3. Architecture/Space — building, interior, city...
4. Food/Lifestyle — dishes, everyday objects...
5. Other — just tell me your idea

Pick a direction, or just give me a topic keyword.

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

Step 6: When user wants to find models themselves

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

Recommended 3D model websites:

1. Sketchfab (largest 3D model platform)
   • Link: https://sketchfab.com/search?q=[keyword]&type=models&downloadable=true
   • Filter steps: search keyword → check "Downloadable" → format "glTF" → sort by "Likes"
   • When downloading, select glTF (.glb) format
   • Note: some models require free registration to download
2. Poly Pizza (all free low-poly)
   • Link: https://poly.pizza/
   • All CC0 licensed — click Download to get .glb directly
   • Best for: minimalist or cartoon-style presentations
3. Sketchfab popular categories
   • Animals: https://sketchfab.com/search?q=animal&type=models&downloadable=true
   • Food: https://sketchfab.com/search?q=food&type=models&downloadable=true
   • Tech: https://sketchfab.com/search?q=gadget&type=models&downloadable=true
   • Architecture: https://sketchfab.com/search?q=architecture&type=models&downloadable=true
4. Free3D (general free model site)
   • Link: https://free3d.com/3d-models/glb
   • Note: check the license type before use
5. TurboSquid Free (pro model site free section)
   • Link: https://www.turbosquid.com/Search/3D-Models/free/glb

After downloading, share the .glb file with me. If the download is a .gltf folder, use Blender to convert it to .glb.

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

Remind about token cost before searching:

I can search for you, but web searches use extra tokens. Would you prefer:

A. I search — I use the Sketchfab API and recommend 2-3 options (uses a few tokens) B. Self-service — I give you search links and filter steps, you pick and share with me (no extra tokens)

A or 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: "Please check the model license before downloading. CC0 / CC BY = free to use; CC BY-NC = non-commercial only."

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
• For additional inspiration, browse ../morph-ppt/reference/styles/INDEX.md — 50+ visual styles organized by mood (dark, light, warm, vivid, bw). Read style.md for design philosophy, build.sh for implementation reference. Learn the approach, do not copy coordinates verbatim

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 fox, duck, avocado): 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.