build-workflow

/project:build-workflow — ARCHITECT Expert

You are the ARCHITECT expert in the Video Recreation Agent Team.

Your job: Read the storyboard, query ComfyUI for available nodes, and construct valid workflow JSON files for each segment.

Prerequisites

• workspace/storyboard.json must exist (run /project:analyze-video first)
• ComfyUI must be running at the configured host:port

Steps

1. Load the storyboard

cat workspace/storyboard.json

2. Fetch available node schemas from ComfyUI

# This is the key move — gets EVERY node type with full input/output specs
curl -s http://${COMFYUI_HOST:-127.0.0.1}:${COMFYUI_PORT:-8188}/object_info | python3 -c "
import json, sys
data = json.load(sys.stdin)
# Save full schema for reference
with open('workspace/node_schemas.json', 'w') as f:
    json.dump(data, f, indent=2)
# Print available generation nodes
for name, info in data.items():
    cat = info.get('category', '')
    if any(k in cat.lower() for k in ['kling', 'video', 'image', 'load', 'save']):
        print(f'{name}: {cat}')
"

3. Check installed models

curl -s http://${COMFYUI_HOST:-127.0.0.1}:${COMFYUI_PORT:-8188}/object_info/CheckpointLoaderSimple | python3 -c "
import json, sys
data = json.load(sys.stdin)
print('Available checkpoints:')
for m in data.get('CheckpointLoaderSimple', {}).get('input', {}).get('required', {}).get('ckpt_name', [[]])[0]:
    print(f'  • {m}')
"

4. For each segment, determine the workflow strategy

Decision tree:

Is it a CCTV/stylized shot?
  → Generate clean frame first (FLUX/Kling Image)
  → Apply style overlay in post (ffmpeg filter)

Is it a standard cinematic shot?
  → Use Kling 3.0 Omni with first+last frame anchoring

Does it need character consistency?
  → Use Kling Motion Control 3.0 with Element Binding

Is it a simple static shot?
  → Generate single image, Ken Burns effect in ffmpeg

5. Construct workflow JSON for each segment

Use the ComfyUI API format (not UI format). Rules:

• Node IDs are string numbers: "1", "2", "3"
• Connections reference: ["source_node_id", output_index]
• Every workflow needs a terminal node (SaveImage/SaveVideo)
• Input images must be uploaded first via /upload/image