Author here. We built Fluxwing to solve a frustrating problem: *design iteration is too slow*.

*The problem:* Figma forces you to start at high fidelity. You're making pixel-perfect decisions before you know if the structure is even right. Want to try 5 different layouts? Hours of dragging and resizing. AI can't help because binary files.

*The workflow we wanted:* 1. Start LOW-fidelity (ASCII) → iterate instantly with AI 2. Add fidelity PROGRESSIVELY as design solidifies 3. Generate HIGH-fidelity outputs (Figma, React) FROM the spec

*The format:* Built on *uxscii* (open standard). Two files: `.uxm` (JSON metadata) + `.md` (ASCII template).

*The magic:* Both humans AND AI understand ASCII natively. No vision models. Instant feedback loop through conversation.

*Progressive fidelity:* - Level 1: ASCII layout (structure) ← start here, iterate fast - Level 2: Component metadata (states, behavior) - Level 3: Design tokens (colors, spacing) - Level 4: Generate outputs (Figma, React)

*The implementation:* Six Claude Code skills. Natural language activation. "Create a login screen" → ASCII in seconds → iterate through conversation → add metadata → done.

*Roadmap:* Building generators that create Figma components and React code FROM .uxm specs. One spec → multiple outputs. Change once, regenerate everything.

*Stack:* Plain text + JSON Schema. No dependencies. Version control friendly. Open source (MIT).

This isn't replacing Figma. It's a *different starting point*. Fluxwing for rapid iteration at low fidelity. Figma for final polish. Or generate Figma FROM Fluxwing when we ship the plugin.

Built by Trabian. Came from internal needs (fintech UIs, lots of iteration with clients).

Happy to answer questions

This is a cool idea!

But my question is: why not just use HTML+CSS?

This is what I do, and it works amazingly well in my experience. LLMs are really good at HTML, and they can generate a clickable prototype in a few seconds.

And the benefit of HTML is that it can actually resemble the thing I'm going after. It's still throw-away code -- when it comes time to build, this HTML is usually used in screenshot form as I might use a quick and dirty Figma mock-up.

Going through ASCII feels like forcing the LLM to go through a translation layer that is going to make things more complicated (like needing 6 skills to be able to do it, for instance).