Roadmap · v2.0 → v3.0 · public · 2026–2027

ux-skill roadmap — v2.0 to v3.0.

Where we are, where we're headed, and what we will never build. Every release has a codename, a ship target, and a vote-on-this link so you can push the order of work. Dates are intent, not promise — this is a solo project with a real day job. If you want something faster, send a PR.

v2.0.0-alpha.24 "Forcing Function"
Shipped · 2026-05-28

Now: the engine, the manifests, the linter, the IDE bridge.

Everything ux-skill is named after. The Python engine reads nine manifests, the 5-parallel-search recommender returns a structured plan, the 22 slash-commands run inside Claude Code / Cursor / Windsurf / 14 other adapters, and the anti-slop linter blocks 35 fingerprints from ever reaching git add.

What's in v2.0

  • 998 manifest entries across anti-patterns, generators, styles, components, brands, foundations, output formats, conditionals, processes
  • 22 slash-commands/ux-design, /ux-component, /ux-system, /ux-audit, /ux-frame, /ux-dashboard, /ux-fix, /ux-copy, /ux-motion, /ux-a11y, and 12 more
  • 17 IDE adapters — Claude Code, Cursor, Windsurf, Copilot, Continue, Codeium, Cline, Roo Code, Aider, Cody, Tabnine, Supermaven, Zed AI, JetBrains AI, Augment, Pieces, Sourcegraph
  • 110 brand specs under references/brands/ — Apple, Stripe, Linear, Vercel, Notion, Airbnb, Figma, Replit, Anthropic, plus 63 more
  • 100 anti-slop linter rules — gradient regex, font-pair lockout, emoji detection, default-blue trap, three-tier pricing card flag, generic hero detection, 29 more
  • Discovery protocol — 10-field intake before any generation (banned: improvisation)
Read the release notes
v2.1 "Stoic"
Planned · Q3 2026

Deepen the linter. Cross-wire the manifests. Add landing-patterns.

v2.0 ships with 35 rules. v2.1 doubles down: 17 more rules pulled directly from the audit of the 50 landing pages I shipped this year that the model defaulted into something generic. Plus a new landing-patterns.yaml manifest with the 28 patterns we keep needing (hero variants, pricing variants, testimonial variants) and cross-reference wiring so a chosen brand auto-narrows the available components.

Key features

  • +17 anti-pattern rules — pulled from the May 2026 audit of 50 generic landings (grep findings: 14 purple gradients, 11 three-tier cards, 9 default-Inter pairs, 16 generic feature-icon grids)
  • landing-patterns.yaml — 28 patterns covering hero, pricing, testimonial, FAQ, CTA, footer variants with brand-locked recommendations
  • Cross-reference wiring — choosing a brand from the 72-pack now auto-narrows valid components, valid type pairings, and valid motion patterns
  • Linter performance pass — drop median run from 380ms to under 80ms; parallel rule eval
  • Per-rule disable file.uxskill.toml with rule-level opt-out and explanation requirement
Vote on v2.1 priorities
v2.2 "Polyglot"
Planned · Q4 2026

Localize the README. Add an MCP server. Bridge to Figma.

Open-design hit 54k stars partly because the README ships in 11 locales. ux-skill is currently English-only and that's a real disadvantage in MENA, China, Japan, Korea, Brazil. v2.2 fixes that. It also adds an MCP server so cross-tool memory (Cursor remembers what Claude Code generated yesterday) is real, and a Figma MCP bridge so design-token import becomes one command.

Key features

  • README in 11 locales — English, Arabic, Spanish, Portuguese, French, German, Russian, Chinese (Simplified), Japanese, Korean, Hindi
  • MCP server (uxskill-mcp) — stateful memory of prior designs across Claude Code, Cursor, Windsurf sessions; canonical brand persists per-repo
  • Figma MCP bridge — pull tokens from a Figma file into a brand spec; one-way sync; design system import in under 30 seconds
  • RTL pass — every command flag-tested under dir="rtl"; mirrored animation contracts in motion-engineer agent
  • Brand pack #2 — 28 MENA + APAC brands added to the brand library (Careem, Talabat, Noon, Aramex, GO-JEK, Grab, Lazada, +21)
Vote on v2.2 priorities
v2.3 "Atelier"
Planned · Q1 2027

Hi-fi prototype output. Local preview server. Motion MP4 export.

huashu-design ships hi-fi prototype output and that's the only dimension where they outscore us. v2.3 closes it. We add a hi-fi mode that emits real component code (React, Vue, Svelte, vanilla) with frame-accurate motion specs, a local preview server on :4242 that watches and re-renders, and MP4 export of every motion spec so you can drop it in a deck.

Key features

  • Hi-fi output mode/ux-design --fidelity hi emits production component code in React, Vue, Svelte, vanilla TS
  • Local preview servernpx uxskill preview opens http://localhost:4242 with hot-reload of every generated artifact
  • Motion MP4 export — every /ux-motion spec exports a 4-second loop MP4 + WebM for decks and docs
  • Component library export — generated systems compile to a publishable npm package on demand
  • Frame-accurate motion engine — sub-16ms GSAP timeline export; cubic-bezier per axis; no ease-in-out defaults
Vote on v2.3 priorities
v3.0 "Mirror"
Planned · Q2 2027

Real-time live preview. Image-to-code reverse bridge. Multi-language gen.

The 1.0 of the next phase. Real-time means the preview server is no longer file-watched; the AI agent streams component diffs into the open browser tab as it generates. Image-to-code closes the loop: paste a screenshot of someone else's UI, get a ux-skill-clean rebuild of it (with the slop detected and removed). Multi-language adds Swift, Kotlin, and Flutter generators so the same brand spec lints native too.

Key features

  • Real-time live preview — agent streams diffs; preview repaints sub-second; no file-watch latency
  • Image-to-code bridge — paste a UI screenshot, get a brand-clean React + Tailwind rebuild with anti-pattern annotations
  • Multi-language code gen — Swift (SwiftUI), Kotlin (Compose), Flutter, Jetpack Compose adapters
  • Linter as language server — LSP integration so anti-pattern violations show inline in any editor that speaks LSP
  • Public APIuxskill score <url> returns a JSON report; bring your own brand
Vote on v3.0 priorities
Won't-do list

What we will never build.

Discipline shows in what you refuse. These are not "not yet" items — they are permanent no's. If a v3.5 ever ships any of these, I owe the community a public retraction.

  • Hosted SaaS The whole point is that the engine runs on your machine. No remote server, no quota, no rate limit.
  • Paid tiers No Pro plan, no premium brand pack, no commercial license. The manifest must stay open.
  • Account system Nothing to sign up for. Ever. The CLI does not authenticate against any service.
  • Telemetry No usage events, no error reporting, no anonymous IDs. Bug reports come from humans via GitHub.
  • AI image generation ux-skill recommends; it does not synthesize. We will not call Midjourney or DALL-E from the CLI.
  • Video generation Same. We may export specs to MP4 (v2.3) but we will never generate video content via a model.
  • Built-in analytics We don't measure user behavior in your repo. The linter reports what's there; that's the only signal.
  • Vendor-specific lock-in No "ux-skill only works in Claude Code." Every adapter ships at parity. Cursor and Windsurf are not second-class.