Kit Layout

Kit layout

One kit per brand, on storage your company approves, outside every repository.

KIT/
  brand.json                 names, triggers, sites, model, ComfyUI URL, picked LoRAs
  .gitignore                 "*"
  auth/<site>/               browser profile from login (secret: it can open the site as you)
  capture/<site>/pages/      full-page screenshots
  capture/<site>/imports/    your screenshots
  capture/<site>/capture.jsonl   one record per picture: url, title, viewport, scale, background, components
  capture/<site>/tokens/     computed-style tallies, custom properties, font faces, component recipes
  tokens.json  brand.css  tokens.md  palette.png
  dataset/<lora>/NNNN.png + .txt + manifest.jsonl      dataset/.captions.json (caption cache)
  review.html
  train/<name>.json .log .pid   train/out/<name>/ (checkpoints, samples)
  loras/<lora>-<model>.safetensors
  out/<time>-<slug>/         mock-ups with mockup.json, or sweep.jpg + sweep.json

brand.json

Field	Meaning
brand.name, brand.trigger	the brand and its made-up trigger word
sites[]	name, trigger, urls, lora (false: feeds the brand LoRA only), blur (selectors blurred on every capture)
model	the training base: zimage, zimage_turbo, qwen_image, klein4b
comfy	the ComfyUI URL for gen, sweep and ref
comfy_loras	LoRA name to ComfyUI file name; pick --to fills it in
comfy_models	per-graph overrides: unet, clip, clip_type, vae, steps, cfg, sampler, scheduler, shift
ai_toolkit	the trainer folder
picked	which checkpoint each LoRA came from, and when

Environment variables

Variable	Default	Used by
BRANDWRIGHT_COMFY	http://127.0.0.1:8188	new kits, doctor
BRANDWRIGHT_COMFY_LORAS	none	pick
BRANDWRIGHT_AI_TOOLKIT	none	train --run, doctor
BRANDWRIGHT_OLLAMA	http://127.0.0.1:11434	caption, doctor
BRANDWRIGHT_VLM	qwen3-vl:8b	caption