Repository navigation
Home
Mark Rogers edited this page Oct 22, 2026
·
7 revisions
brandwright makes mock-ups of new UI that look like they came from your own sites. You point it at your brand's websites, or hand it screenshots. It learns the look three ways and gives you mock-ups of components and pages that do not exist yet.
| Route | What you get | Needs | Time |
|---|---|---|---|
| Code | an HTML page on brand.css (the site's real colours, fonts, variables and button, input and card styles), rendered to PNG; exact text |
a browser | minutes |
| Reference | a picture in the style of up to four screenshots, made with FLUX.2 klein | ComfyUI with klein 4B | about 30 s a picture |
| LoRA | a brand LoRA plus a LoRA per sub-site that learns only what the site adds; on-brand pictures from any prompt | an NVIDIA GPU, 16 GB | hours once, then seconds |
It is an agent plugin, with four skills for Claude Code and other agents that read SKILL.md, and one Python CLI, bw.py, that also works on its own.
- Getting started: install, first kit, first capture, first code mock-up.
- Routes: which route for which job, and how they combine.
-
Capture and tokens: crawling, logging in, importing screenshots, masks,
brand.css. - Dataset and captions: tiles, component specimens, dedupe, caption rules.
- Training: base models and licences, residual and continue modes, checkpoints.
-
Mock-ups: code route,
gen,sweep,ref, scoring, prompt patterns. - Commands: every command and option.
-
Kit layout: the folder and
brand.json. - Myths and facts: the usual brand-LoRA advice, checked.
- Recipes: common jobs, start to finish.
- FAQ.
- Development: tests, budget, releases.
Version 0.1.5, released 2026-10-22 (release notes). Source: m4bwav/brandwright.
brandwright 0.1.5 · repository · issues · MIT