Claw3d Print Workflow

name: claw3d description: "Unified 3D workflow: create models (AI), search (Thingiverse), slice, print. Modular—enable only what you need." metadata: { "openclaw": { "emoji": "🖨️", "skillKey": "claw3d", "primaryEnv": "FAL_API_KEY", "requires": { "anyBins": ["claw3d"], "env": ["FAL_API_KEY","THINGIVERSE_ACCESS_TOKEN"] }, "homepage": "https://github.com/makermate/openclaw-3d", "install": [ { "id": "uv", "kind": "uv", "package": "claw3d", "bins": ["claw3d"], "label": "Install claw3d (via uv)" } ], "setupNote": "FAL_API_KEY: required for 3D generation (use input above or .env). GEMINI_API_KEY: optional but recommended — enables image intent analysis and video-to-3D (Gemini 2.5 Flash, free tier at aistudio.google.com/apikey). THINGIVERSE_ACCESS_TOKEN: add to .env for model search. VIDEO SUPPORT: OpenClaw defaults to a 5MB media limit per channel. To allow videos up to 50MB via Telegram, add to openclaw.json under channels.telegram: {"mediaMaxMb":50}" }

}

claw3d — Unified 3D Workflow

Single skill for the full 3D pipeline: create models (AI), search (Thingiverse), slice, and print. Modular—only enabled capabilities appear below.

CRITICAL — Execute immediately, never stop after acknowledging. When you tell the user you'll do something (search, convert, preview, slice), you MUST execute the command in the SAME turn. Do NOT end your turn after just acknowledging — the user should never have to ask "done?" to get you to act. Acknowledge AND call the tool in one response.

Setup order: Printer first (when printing enabled) → then create/search models → slice → print.

CRITICAL: Setup Flow — Printer + Profile First

When printing is enabled and the user has no printers configured, guide them to add a printer AND a linked slicing profile before creating or searching for models. A linked profile is required for slicing — it stores the build volume (width × depth × height) extracted from the 3MF, which determines how models are scaled.

Always run claw3d printer list first. If it returns nothing, go through setup below.

Step 1 — Ask for printer info + 3MF

Send this message to the user:

Let's get your printer set up. I need 3 things:

1. Printer name — e.g. "Creality K2 Pro Living Room"
2. IP address + port — e.g. 192.168.1.50:7125 (Moonraker default: 7125; Creality K2 SE: 4408)
3. Cura project file (.3mf) — Export it from Cura: File → Save → "Export Universal Cura Project" with your printer loaded. This file carries your printer's build volume and all settings — it's required for correct slicing.

Wait for the user to provide all three.

Step 2 — Add printer with profile in one command

claw3d printer add --name "<name>" --host <ip> --port <port> --profile-from-3mf <MediaPath>

This does everything in one step:

• Registers the printer (name, IP, port) in ~/.config/claw3d/config.json
• Extracts the Cura machine + extruder definitions from the 3MF
• Creates a slicing profile on the slicer server (stores build_width, build_depth, build_height)
• Links the profile to the printer as its default

If the user provides name+IP but no 3MF yet: Add without it (printer add --name ... --host ... --port ...), then immediately ask for the 3MF to create the profile:

Got it! Now please send the Cura project file (.3mf) so I can create the slicing profile. In Cura: File → Save → "Export Universal Cura Project".

Then: claw3d profile create --from-3mf <MediaPath> --name "<printer_name>_profile" → claw3d printer set-profile <printer_id> <profile_id>

Printer backends: Run claw3d configure backends to see options (Moonraker, PrusaLink, etc.). Community can add backends in claw3d/backends/.

CRITICAL: When User Asks for a Model (Vague Request)

When the user asks for a 3D model without specifying how (e.g. "I need a cup", "I want a dragon", "find me a vase", no image attached), do NOT default to one option. Offer choices based on what's enabled:

Great! Would you like me to:

1. Search for existing models — I'll look on Thingiverse and show you options to download (if directory enabled)
2. Create a 3D model from an image — Send me a sketch or photo and I'll turn it into 3D (if ai-forger enabled)
3. Search first, then create from an image if nothing fits — Best of both (if both enabled)

Wait for the user to choose. Only if they explicitly say "create it", "from a photo/sketch", "search", "look up", etc., then proceed.

Never assume — "I need a dragon" could mean search OR create from image. Always clarify when ambiguous. Do not offer text-only 3D generation — results are inaccurate; always require an image or sketch.

Shared Rules: MediaPath and Unique IDs

MediaPath: When the user attaches a file (image, GLB, 3MF), the message includes a MediaPath — the full filesystem path. Always pass that exact path to --image, --edit-3d, --profile-from-3mf, etc. Copy it character-for-character.

Unique output paths: The workspace is shared. Using fixed names (model.glb, preview.mp4) causes old files from a previous request to be sent to new chats. Always derive a short ID from the MediaPath and use it for outputs.

MediaPath format: .../file_13---b10560d7-18fd-40e9-8a49-996ad190a26c.jpg — extract the segment after --- and use the first 8 chars (e.g. b10560d7) as ID.

If the MediaPath has no UUID (unusual), use date +%s to get a unique ID.

User Sends an IMAGE and Asks to 3D Print

When the user attaches an image and asks to "3D print this", "print this", "make it printable", etc. — you CAN do it (if ai-forger + slicing + printing enabled):

1. Acknowledge first — "Yes! I'll turn that into a 3D model and get it ready to print. Give me a minute."
2. Convert — claw3d convert --image <MediaPath> --output model_<ID>.glb
3. Get build volume — Run claw3d printer list; note [WxDxH mm] if shown.
4. Preview — claw3d preview --input model_<ID>.glb --output preview_<ID>.mp4 [--build-volume WxDxH] — send the video
5. Slice — Run claw3d profile list, then slice with --build-volume <WxDxH> and profile or --profile-from-3mf
6. Print — Run claw3d printers, then claw3d print --gcode model_<ID>.gcode

Do NOT say "I can't print from an image" — you can create the 3D model first. If FAL_API_KEY is missing, convert will fail; then tell the user to set it up.

Workflow Overview

Get model (search OR create) → optionally edit → slice → print

• Search — claw3d search → claw3d fetch → claw3d dimensions → present with preview
• Create — claw3d convert --image (requires image/sketch) → claw3d preview → present
• Edit — claw3d convert --edit-3d (when user sends GLB and asks to modify)
• Slice — claw3d slice (sends G-code + gcode preview video)
• Print — claw3d print

Commands (Overview)

Run all via exec. Use claw3d.

CRITICAL: Never Expose Internal Reasoning to the User

All routing decisions, skill logic, and internal reasoning are for YOUR use only. NEVER send them to the user. The user should only see friendly, concise messages — never references to "Primary Gate", "SKILL.md", module names, decision rules, or your thought process. If you need to reason about which path to take, do it silently. The user just wants their model.

Bad (leaked reasoning): "According to the Primary Gate, a wine stand is a common functional object, so I should search Thingiverse..." Good (user-facing): "Great, let me take a look at what you need — give me a moment!"

CRITICAL: Primary Gate — Search or Create?

This is the FIRST decision for EVERY request — images, videos, and text. Run the Primary Gate BEFORE any analysis, frame extraction, or claw3d analyze. Make this decision silently — do NOT explain your routing to the user.

The key question is: Would an existing Thingiverse model likely satisfy this need, or does the request require something inherently custom/unique?

Primary Gate: SEARCH vs CREATE

→ SEARCH path first (go to 03-directory module) when:

• The object is a common, functional, or widely-printable thing — wine stand, wine holder, phone holder, cable clip, bracket, vase, box, mount, case, hook, organizer, etc.
• Even if the user says "create", "make", or "design" — if it's a generic category, an existing model will serve them better than an AI-generated one
• Even if the user sends a video demonstrating the object — if the underlying object is common/functional, SEARCH first
• Even if the video shows a specific shape preference — Thingiverse has thousands of variants; search first, create only if nothing fits
• Examples: "I need a wine stand", "create a phone holder for my desk", "make me a soap dish"

→ CREATE path (continue to CREATE section below) when:

• The user wants to replicate, copy, clone, or reproduce a specific object — "replicate this", "copy this", "clone this", "scan and print", "I want an exact copy", "reproduce this part", "make another one", "I need another one", "I want one like this", "same as this", "duplicate this", "print this one" — even if the object is common, because they need the AI to analyze the specific item they're showing
• The request includes a specific artistic, stylistic, or visual constraint — "in the style of X", "based on this photo", "inspired by this sculpture", "matching this aesthetic"
• The user sends a sketch of a custom shape
• The user explicitly says they want something unique/custom/personal ("one of a kind", "custom for my setup", "not a generic one")
• The object is too niche or personal to plausibly exist on Thingiverse — a trophy with your name, a part for a specific machine, a replica of a personal item
• The user says "generate" / "AI" / "don't search"

Decision rule of thumb:

"Could I type this into Thingiverse and find 5+ decent results?" → YES → SEARCH first "Does this require seeing a specific image, style, or personal constraint to design?" → YES → CREATE

→ ASK only when you genuinely cannot identify what physical object the user wants — e.g. "make something for my office" with no further context. If you can name the object, go to SEARCH. Do not ask.

SEARCH PATH — fallback to CREATE: After 3 rounds of search (up to 15 models reviewed) with no match, tell the user nothing matched and ask if they want a custom AI-generated model instead. If they have a photo/video, use it as reference for the AI generation.

Applying the Primary Gate to Videos

When the user sends a video, you may receive a text Description (from OpenClaw's Gemini video understanding). Use the Description and/or the user's message text to run the Primary Gate — BEFORE extracting any frame or running claw3d analyze.

Steps for video:

1. Read the user's message text + any Description
2. Identify what physical object they want (e.g. "wine holder", "phone stand", "bracket")
3. Run the Primary Gate on that object name
4. If SEARCH → go directly to 03-directory module with that object as the search query. Do NOT extract a frame or run analyze
5. If CREATE → continue to the "When User Sends a VIDEO (CREATE path)" section below

⚠️ CRITICAL: A video showing someone demonstrating a common object does NOT make it a CREATE request. The video is just their way of communicating what they want — it doesn't mean they need AI generation. A person holding up a wine bottle and showing how they'd like a wine stand still maps to SEARCH. Only explicit artistic/stylistic/replication intent maps to CREATE.

When User Sends a VIDEO — Get the File

This section handles finding the video file. It applies to BOTH paths (the CREATE path needs it for frame extraction; the SEARCH path may need it later if search fails and you fall back to CREATE).

Step 0 — Acknowledge immediately — Before doing anything else, send:

"Great, let me take a look at what you need — give me a moment!"

Step 1 — Get the video path. Three cases:

Case A — File path visible ([media attached: /home/node/.openclaw/media/inbound/...]): Use that exact path.

Case B — No file path but Description is present (OpenClaw's Gemini video understanding ran and suppressed the path): The video is still on disk. Find it:

ls -t /home/node/.openclaw/media/inbound/ 2>/dev/null | head -5

Pick the most recent video file (.mp4, .mov, .webm). Use that as the path.

Case C — No file path and no Description (video silently dropped — too large):

Your video was too large — OpenClaw's default limit is 5MB. I can increase it to 50MB right now. Want me to?

If confirmed:

python3 -c "
import json, pathlib
p = pathlib.Path('/home/node/.openclaw/openclaw.json')
cfg = json.loads(p.read_text())
cfg.setdefault('channels', {}).setdefault('telegram', {})['mediaMaxMb'] = 50
p.write_text(json.dumps(cfg, indent=2))
print('Done')
"

Reply: "Done! The limit is now 50MB — please resend your video."

The config watcher restarts the Telegram channel automatically.

Step 2 — Run the Primary Gate using the Description/user message → SEARCH or CREATE. See above.

If SEARCH → go to 03-directory module. Note the video path — if search fails and you fall back to CREATE, you'll need it for frame extraction.

If CREATE → continue to the next section.

When User Reports "File too large" / Video Rejected

The bot rejects oversized files before the agent sees them. If the user reports this error in a text message, offer to fix it:

I can increase your video limit to 50MB right now. Want me to do that?

If confirmed, run the patch above.

Analysis Modes

Run once per session to understand the configuration:

claw3d configure analysis --status

CREATE Path: Intent Analysis for Image or Video

Only enter this section if the Primary Gate resolved to CREATE.

Before doing anything with a user's image or video, run claw3d analyze (images) or analyze the video natively + claw3d extract-frame --timestamp + claw3d analyze (videos).

When User Sends an IMAGE (CREATE path)

Step 1 — Always run analyze:

claw3d analyze --input <MediaPath> [--description "user's message"] [--pretty]

Step 2 — Read the result and branch:

Result has native_mode: true → you are the analysis layer

Analyze the image yourself using these rules:

Classify image_type:

• sketch: hand-drawn, pencil/pen outlines, whiteboard drawings → intent is almost always create_new, proceed directly
• photo: real photograph → read description carefully
• 3d_model: CAD rendering or existing 3D model screenshot
• reference: product photo, inspiration, logo

Decide needs_clarification:

OVERRIDE — replicate/copy intent always sets needs_clarification: false: If the user's message contains any of: "make another one", "copy this", "replicate this", "clone this", "I want one like this", "same as this", "reproduce this", "duplicate this", "print this one" — the photo/frame IS the complete design reference. Proceed directly to convert. Do NOT ask for a drawing. The whole point is that they're showing you the exact object they want.

false (proceed without asking) when ALL of these are true:

• Single clear subject identified
• The description or image already specifies at least one key design constraint (size, mounting type, orientation, number of units, etc.)
• Sketches always qualify — the drawing itself conveys the shape intent

true (ask ONE clarifying question) when ANY of these:

• Complex scene with multiple objects and no description
• Custom functional or structural object (holder, bracket, stand, organizer, case, clip, mount, etc.) where the description does NOT specify key design details — even if the object is clearly identified — but only if there is no replicate/copy intent (see override above)
• Subject is clear but could be made many ways (e.g. "a wine holder" — wall-mount or freestanding? holds 1 bottle or multiple? specific angle?)
• Abstract or landscape photo with no description
• "Make this better" / "improve this" with no context

Rule of thumb for functional objects: If you could design it 3+ different ways and the user hasn't said which way → send the frame/image back and ask them to draw on it (see below). Exception: replicate intent (see override above) → always proceed directly.

If needs_clarification: false:

Step A — Tell the user you're starting (do NOT stay silent):

"Creating your 3D model now — I'll send it when it's ready!"

Step B — Write a suggested_prompt and run convert:

claw3d convert --image <MediaPath> --prompt "<suggested_prompt>" --output model_<ID>.glb

CRITICAL — When writing suggested_prompt:

For replicate/copy intent ("make another one", "copy this", etc.): Keep it SHORT — one sentence. The image already carries the shape. Do NOT add dimensions, material suggestions, or printing advice.

• ✅ "a replica of the black S-hook, matching its exact shape for hanging kitchen utensils"
• ✅ "a replica of the wooden phone stand shown in the image"
• ❌ "A 3D model of a sturdy S-shaped utility hook, designed for 3D printing, with a flat bar profile and rounded edges. The hook should be approximately 7-8 cm in length..." ← WAY too long, invents dimensions

For all other intents: Describe ONLY the 3D object to be printed. Keep it to 1-2 sentences max. Do NOT include:

• Dimensions or measurements (the image conveys scale)
• Material or printing recommendations (PETG, PLA, etc.)
• Scale references ("sized based on the dog for scale")
• People, hands, or human body parts visible in the image
• Background items, decorations, scene context

Example — user shows a wine bottle next to a dog sculpture:

• ❌ WRONG: "An L-shaped wine holder sized appropriately based on the teal dog sculpture for scale"
• ✅ RIGHT: "An L-shaped wine bottle holder with a circular opening at a 45° angle, wall-mountable"

If needs_clarification: true:

Two cases:

Case 1 — Ambiguous subject (multiple objects, unclear what to print): Ask ONE specific text question:

"I see a desk with a laptop and a mug. Which item would you like to 3D print?"

Case 2 — Subject is clear but it's a photo of a functional/custom object (holder, bracket, case, mount, stand, organizer, etc.):

1. Note the original frame path (e.g. frame_1a589237.jpg) — you will need it when the annotated image comes back.
2. Send the extracted frame back to the user and ask them to draw on it in red:

   "Hey! Could you draw in red on this image to show me the shape you have in mind? Any drawing app works — even a quick scribble on your phone. Then send it back and I'll use it as the design reference." Use the message tool to attach the frame — do NOT use inline MEDIA: syntax: message(text="Hey! Could you draw...", media="<frame_path>")

3. Wait for the user to send back the annotated image.

When the user sends back the annotated image: Do NOT say "Yes! On it!" and stop — immediately run exec:

claw3d convert --image <original_frame_path> --annotated-image <annotated_MediaPath> --prompt "<description of the object, NO scene context>" --output model_<ID>.glb

• <original_frame_path> = the frame you sent them (e.g. frame_1a589237.jpg)
• <annotated_MediaPath> = the absolute path from the media attached message
• Then run preview + send both files as usual

Result has native_mode: false (Gemini was used) → act on the JSON

{
  "subject": "a wooden phone stand",
  "image_type": "sketch",
  "intent": "create_new",
  "needs_clarification": false,
  "clarification_question": null,
  "suggested_prompt": "a minimalist wooden phone stand with a 70° angled back support..."
}

If needs_clarification: true:

• First check for replicate/copy intent in the user's description — if present ("make another one", "copy this", "replicate", etc.), override to false and proceed directly regardless of what Gemini returned.
• Otherwise: send clarification_question verbatim (Gemini wrote it to be friendly and specific)
• Do NOT rephrase
• After user replies, re-run: claw3d analyze --input <MediaPath> --description "<original + reply>"
• After one round, always proceed

When User Sends a VIDEO (CREATE path)

You should only be here if the Primary Gate resolved to CREATE.

Step 1 — Extract the best frame

Two paths depending on how the video arrived:

Case A — Video attached as media (you can see the video in this conversation): You are a multimodal agent. Analyze the video directly to identify the best frame:

• Subject fully in frame, clear and well-lit
• Best reveals the 3D shape (front 3/4 angle preferred)
• Not blurry, not mid-motion, not transitioning

Pick the exact timestamp (HH:MM:SS), then extract:

claw3d extract-frame --input <video_path> --timestamp <HH:MM:SS> --output frame_<ID>.jpg

Case B — Only text Description, no media in conversation (OpenClaw pre-processed the video): You cannot see the video — you only have the text Description. Do NOT guess a timestamp from text. Use Gemini API for smart frame selection:

claw3d extract-frame --input <video_path> --output frame_<ID>.jpg

(no --timestamp → Gemini picks the best frame automatically)

If this fails because no Gemini API key is configured, stop and tell the user:

"I need a Gemini API key to pick the best frame from your video (the video isn't directly visible to me in this conversation). Please run: claw3d configure analysis --gemini-api-key <YOUR_KEY> You can get a free key at Google AI Studio."

Step 2 — Analyze extracted frame:

claw3d analyze --input frame_<ID>.jpg --description "<user's message or Gemini description>" --pretty

Then follow the IMAGE flow above (including needs_clarification checks).

CRITICAL — Do NOT go silent after frame extraction. If needs_clarification: false, tell the user you're generating the model BEFORE running claw3d convert. The full sequence must be:

1. "Great, let me take a look at what you need — give me a moment!" (from earlier)
2. Extract frame + analyze (Steps 1-2)
3. "Creating your 3D model now — I'll send it when it's ready!" ← say this BEFORE convert
4. Run claw3d convert → claw3d preview → send both files

Full Example Flows

Generic functional object — search first (even if user says "create"):

User: [sends video] "I need you to create a wine stand"
→ Primary Gate: wine stand = common, functional → SEARCH path
→ Go to 03-directory module: search → thumbnails → pick → confirm → download → preview

Video demonstrating a common object — STILL search first:

User: [sends video showing how they'd hold a wine bottle, describing an L-shaped holder]
→ Primary Gate: wine holder = common, functional → SEARCH path (video demo ≠ custom design)
→ Go to 03-directory: search "L-shaped wine bottle holder" → thumbnails → pick

Same object + artistic constraint — create:

User: [sends video + photo of sculpture] "I need a wine stand in the style of this sculpture"
→ Primary Gate: style constraint present → CREATE path
→ claw3d extract-frame → analyze (photo as reference) → convert with prompt + image

Sketch → 3D model (CREATE path):

User: [sends pencil sketch of a bracket]
→ Primary Gate: sketch present → CREATE path
→ claw3d analyze --input sketch.jpg --description "make this"
  (native: sketch → create_new, needs_clarification: false, proceed directly)
→ claw3d convert --image sketch.jpg --prompt "an L-shaped bracket with two mounting holes" --output model_abc.glb

Generic object, user says "I want this" with a photo:

User: [sends photo of a mug] "I want this"
→ Primary Gate: mug = common object → SEARCH path
→ Go to 03-directory: search "mug" → thumbnails → confirm → download → preview

Custom functional object (specific, unlikely to exist):

User: [sends photo of a weird desk edge] "make a phone holder that clips onto this exact edge"
→ Primary Gate: too specific/personal to exist → CREATE path
→ analyze → needs_clarification → ask for sketch on the photo

Video — user asking to find (any wording):

User: [sends video, description: "person asking to find/create a wine stand, demonstrates with bottle"]
→ Primary Gate: wine stand = common functional object → SEARCH path
→ Go to 03-directory module

Search exhausted → fallback to CREATE:

→ 3 searches, 15 thumbnails reviewed, none match
→ "Couldn't find a good match. Want me to generate a custom one with AI?"
User: "yes"
→ If user has a video/photo: use it as reference for CREATE path
→ Extract frame (if video) → analyze → clarification if needed → convert

Commands Reference

# Image intent analysis (outputs JSON)
claw3d analyze --input <image> [--description "text"] [--annotated <image>] [--pretty]

# Video frame extraction
claw3d extract-frame --input <video> [--output frame.jpg] [--timestamp HH:MM:SS]

# Analysis layer configuration
claw3d configure analysis                                 # show status
claw3d configure analysis --mode native                  # use your own AI
claw3d configure analysis --mode auto                    # gemini if available, else native
claw3d configure analysis --mode gemini                  # always use Gemini
claw3d configure analysis --gemini-api-key <KEY>         # set Gemini key
claw3d configure analysis --clear                        # remove Gemini key

Module: AI Model Creation

You CAN convert 2D images to 3D models AND edit existing 3D models (GLB). When the user sends a GLB and asks to modify it (e.g. "make it blue", "add wheels"), use claw3d convert --edit-3d <GLB_MediaPath>. When they send an image, use --image.

When to Use

• "Modify this 3D model" / "Change the color" / "Make it blue" (user sent a GLB) — use --edit-3d flow
• "Turn this sketch into a 3D model" / "Make this image a 3D model"
• "Convert this to 3D" / "Create a 3D model from this photo"
• "Could we 3D print this?" (with image) — convert first, then slice and print
• "3D print this image" / "Print this as a 3D"

Always require an image or sketch. Do not use text-only --prompt for new model creation — results are not accurate enough. If the user asks to "make a cup" without an image, ask them to send a sketch or photo, or offer to search Thingiverse instead.

CRITICAL — User sent a GLB and wants to edit it: Run claw3d convert --edit-3d <GLB_MediaPath> --prompt "..." --output edited_<ID>.glb. Never say you cannot edit 3D models.

Acknowledge Before Converting

When the user asks to create a 3D model from an image OR to edit a 3D model, REPLY IMMEDIATELY first:

• "Yes! Give me a minute—I'll let you know when the 3D model is ready."
• "On it! I'll work on that now and get back to you when it's done."

Image conversion takes 1–2 minutes. Edit-3d can take 5–10+ minutes when Hunyuan is cold. Do NOT stay silent—always acknowledge first.

Critical: Image from Chat

When the user attaches an image, the message includes a MediaPath. Always pass that exact path to --image. Copy it character-for-character.

IMPORTANT — Unique output paths: Derive a short ID from the MediaPath. Format: .../file_13---b10560d7-18fd-40e9-8a49-996ad190a26c.jpg — use first 8 chars after --- (e.g. b10560d7) as ID.

Build volume: Before running claw3d preview, check if a printer is configured with a known build volume: run claw3d printer list and look for [WxDxH mm] (e.g. [350×350×350mm]). If found, pass --build-volume WxDxH (e.g. --build-volume 350x350x350) — this renders the grey build plate and grid under the model for a realistic preview.

ID=b10560d7   # from MediaPath
claw3d convert --image <MediaPath> --output model_${ID}.glb
# With printer build volume (preferred):
claw3d preview --input model_${ID}.glb --output preview_${ID}.mp4 --build-volume 350x350x350
# Without printer configured:
claw3d preview --input model_${ID}.glb --output preview_${ID}.mp4

If the MediaPath has no UUID (unusual), use date +%s for a unique ID. NEVER send model.glb or preview.mp4 that existed before this request.

Edit-3D Workflow (MUST follow in order)

1. Acknowledge — "On it! Editing the 3D model. The Hunyuan step can take a few minutes when cold."
2. Run convert — claw3d convert --edit-3d <GLB_MediaPath> --prompt "..." --output edited_<ID>.glb
   • Edit-3D can take 5–10+ minutes. If it backgrounds, call process poll <session> with timeout: 120000. You will be notified when it completes — do NOT poll in a rapid loop.
   • When you see Wrote edited_<ID>.glb → convert is done.
3. Run preview — claw3d preview --input edited_<ID>.glb --output preview_edited_<ID>.mp4 [--build-volume WxDxH]
   • NEVER use --real-scale for edited models. AI-regenerated models use normalized units (~1 unit), not mm. The preview auto-scales the model to fill the build volume.
   • Same: wait for result or single poll with timeout: 120000.
4. Send BOTH files — TWO message() calls required. You are NOT done after sending the preview.

   message(action="send", text="Here's the updated preview!", media="preview_edited_<ID>.mp4")
   message(action="send", text="And the edited model:", media="edited_<ID>.glb")
   

   CRITICAL: Do NOT end your turn after the first message(). You MUST send the .glb in a second message() call. The user needs the 3D model file, not just the video. Your turn is only complete after BOTH files are sent.

NEVER use --image for a GLB when modifying. --image is for 2D sketches/photos.

Convert Workflow (image/sketch → 3D)

1. Acknowledge — "Yes! Give me a minute—I'll let you know when the 3D model is ready."

2. Check build volume — Run claw3d printer list; note [WxDxH mm] if present.

3. Run convert — claw3d convert --image <MediaPath> --output model_<ID>.glb

   • Convert takes 1–2 minutes. The exec call will wait for it to finish (up to 2 min).
   • If it backgrounds (Command still running), call process poll <session> once with timeout: 120000. You will get notified when it completes — do NOT poll in a rapid loop.
   • When you see Wrote model_<ID>.glb → convert is done. Proceed immediately.

4. Run preview — claw3d preview --input model_<ID>.glb --output preview_<ID>.mp4 [--build-volume WxDxH]

   • Same as above: wait for the result or poll once with timeout: 120000.
   • When you see Wrote preview_<ID>.mp4 → preview is done. Proceed immediately.

5. Send BOTH files — TWO message() calls required. You are NOT done after sending the preview.

   message(action="send", text="Here's your 3D model preview!", media="preview_<ID>.mp4")
   message(action="send", text="And the 3D model file:", media="model_<ID>.glb")
   

   CRITICAL: Do NOT end your turn after the first message(). You MUST send the .glb in a second message() call. The user needs the 3D model file, not just the video. Your turn is only complete after BOTH files are sent.

6. ALWAYS ask about printing — After sending the preview and model, ask:

   Want me to slice this for 3D printing? If so, I need:

   1. Max print size — What's the longest dimension? (e.g. 100mm, 150mm)
   2. Strength — How strong? (10%, 25%, 50%, 75%, or 100%)
   3. Detail — How much print quality? (10%, 25%, 50%, 75%, or 100%)

   This is mandatory for AI-generated models — they have no real-world dimensions, so you MUST get the max print size from the user before slicing. Do NOT slice without asking. Do NOT use a default size.

CRITICAL: Run convert and preview via exec BEFORE sending. The files do not exist until you create them.

Commands

# Image/sketch → 3D
claw3d convert --image <MediaPath> [--prompt "extra description"] --output model_<ID>.glb

# Multiview: 4-quadrant image → Gemini → Hunyuan3D
claw3d convert --image multiview.png --multiview [--prompt "add wheels"] --output model_<ID>.glb

# Edit 3D: modify existing GLB
claw3d convert --edit-3d model.glb --prompt "make it blue" --output edited.glb

# Preview
claw3d preview --input model_<ID>.glb --output preview_<ID>.mp4

# Scale
claw3d scale --input model.glb --output scaled.glb --scale 0.5

Prerequisites

• FAL_API_KEY — Required for convert (set in Control UI → Skills → claw3d)
• Preview — Headless: xvfb and xauth; or use openclaw:claw3d Docker image

Error Handling

API key errors: When convert fails with "FAL API key" or "401/403", ask the user to verify their API key in Control UI → Skills → claw3d. Get a key at https://fal.ai/dashboard/keys

Module: 3D Model Search (Thingiverse Library)

You CAN search for existing 3D models on Thingiverse. Use claw3d find, claw3d fetch, and claw3d preview — do NOT use web search.

When to Trigger This Flow

The intent classifier (06-intent.md) sets consider_search: true for common/functional objects. When that happens — or when the user explicitly asks to find/search/download a 3D model — follow this workflow.

Skip this flow and go straight to AI generation only when the user explicitly says "custom", "artistic", or describes something that clearly doesn't exist as a standard printable part.

Full Workflow: Find → Pick → Preview

Step 0 — Acknowledge (if triggered by video or image)

Before running anything, send:

"Great, let me take a look at what you need — give me a moment!"

Step 1 — Find fitting models (single command)

Run one command. It searches Thingiverse, downloads all thumbnails, fetches each model, packs multi-part models, and fit-checks all of them — returning only the models that physically fit in the printer.

claw3d find "<query>" --max-passing 4

⚠ claw3d find is a long-running command (1–2 minutes). When you see Command still running, this is the SEARCH running — NOT slicing. Do NOT say "Slicing started". Just poll silently with process poll <session> every 10 seconds until it finishes.

Output format:

[1] Balancing Wine Holder
    ID: 660698
    URL: https://www.thingiverse.com/thing:660698
    By: Tanacota
    Thumbnail: thumb_660698.jpg
    Model: model_660698.glb
    Extents: 153.8×160.5×69.8mm
    Rotation applied: none

[2] ...

• Build volume is auto-read from the default printer. Pass --build WxDxH to override.
• Exit 0 = at least one fitting model found. Exit 1 = none found — refine query and retry.
• All model files are already downloaded and pre-rotated (if a rotation was needed to fit).

Step 2 — View thumbnails, pick best 4

View all returned thumbnail files visually (you are multimodal). Pick the 4 that best match what the user needs — considering shape, function, and apparent quality. Record the thing ID for each.

If fewer than 4 passed, use however many there are. If exit 1 (none fit): run claw3d find with a refined query (change keywords, add constraints), up to 3 rounds total. After 3 rounds with no match, fall back to AI generation.

Step 3 — Stamp A/B/C/D badges onto thumbnails, compose grid, send to user

Stamp option letters and compose a single 2×2 grid image (A top-left, B top-right, C bottom-left, D bottom-right):

claw3d stamp-thumbnails --grid thumb_660698.jpg thumb_123456.jpg thumb_789012.jpg thumb_456789.jpg
# Outputs: thumb_660698_A.jpg, thumb_123456_B.jpg, thumb_789012_C.jpg, thumb_456789_D.jpg
# Grid:   thumb_grid_thumb_660698.jpg  (single 2x2 image)

MANDATORY: Send ONE message with the grid image attached. Do NOT describe the options without the image. Do NOT skip the media= parameter. The user needs to SEE the thumbnails to choose — text-only is useless.

message(action="send", text="Here are four options I found:\n\nA — [name/brief reason]\nB — [name/brief reason]\nC — [name/brief reason]\nD — [name/brief reason]\n\nReply with A, B, C, or D — or let me know if none look right and I'll search again or create a custom one.", media="thumb_grid_thumb_660698.jpg")

The media= field with the grid image path is REQUIRED. If you send this message without the grid image attached, the user cannot see the options.

Wait for user response before continuing.

• User picks A, B, C, or D: read the thing ID from the claw3d find output for that option. Continue to Step 4 with that ID. The model is already downloaded and pre-rotated (if needed) — no re-fetch required.
• User says none / not quite: run claw3d find with a refined query, repeat from Step 1 — up to 3 rounds total
• After 3 rounds with no match: fall back to AI generation

Step 4 — Get grouped file list (deterministic extension selection)

The chosen model is already downloaded. Use its thing ID to inspect variant/part structure:

claw3d fetch --list-grouped <thing_id>

This deterministically selects the best extension group (STL > OBJ > GLB > 3MF). Parse the output:

• Best extension: .stl (N file(s)) — what will be downloaded
• Sub-variants (size/version choices…) — only shown if multiple size/version options exist
• Multi-part model (N components…) — only shown if multiple parts with no size variants

Step 5 — Handle sub-variants or cosmetic variations (if any)

Read the --list-grouped output and follow exactly one of these branches:

→ Output shows "Sub-variants" (size/version choices): Ask the user which size/version they want:

"This model comes in several sizes: small, medium, large. Which would work best for you?" Once user picks, use --choose "<variant_tag>" in Step 6.

→ Output shows "Cosmetic variations" (same model, minor differences): Do NOT ask the user. Auto-select the variant marked <- auto-selected and inform them:

"This model has

…(truncated)