Design direction for AI coding agents—grounded in product behavior, platform idiom, and rendered evidence.
Stark routes UI work across web, Windows, Apple, Android, cross-platform, UX, and design tokens. It makes the agent choose the product object, user flow, visual language, states, responsive behavior, motion ownership, and proof strategy before generating code.
Most weak generated interfaces fail before the first CSS rule: they choose a generic structure, decorate it, and call the result polished. Stark changes that sequence.
- Route the request to the correct platform and surface.
- Load a small, relevant reference bundle.
- Write a compact decision brief.
- Build in the repository's existing stack.
- Inspect rendered evidence.
- Repair the highest-impact failure and re-check it.
The runtime uses two core references and at most two conditional references per task. The full library remains available as a catalog without flooding the agent's context.
Install this repository as a plugin, then invoke Stark naturally:
Use Stark to redesign this dashboard around incident triage.
The Codex manifest is at .codex-plugin/plugin.json.
Add the repository as a plugin marketplace source:
/plugin marketplace add f0d010c/stark
/plugin install stark@stark
The compatibility manifest is at .claude-plugin/plugin.json.
Use Stark to design a React analytics dashboard for returning operators.
Prioritize scan speed, realistic states, keyboard use, and mobile replacement.
Use Stark to redesign this SwiftUI onboarding flow.
Preserve native conventions and prove recovery from denied permissions.
Use Stark to audit this landing page for generic structure, weak proof,
typography, accessibility, responsive behavior, and unnecessary libraries.
| Area | What Stark decides |
|---|---|
| UX | Flows, information architecture, first value, states, recovery, repeated-use ergonomics |
| Web | Surface type, stack, composition, typography, copy, assets, responsive behavior, motion |
| Native | Windows, Apple, and Android idioms, controls, navigation, density, materials, accessibility |
| Cross-platform | Shared versus platform-owned behavior for Electron, Tauri, React Native, Flutter, and CMP |
| Design systems | Tokens, component APIs, variants, state galleries, themes, density, drift checks |
| Audit and QA | Usability, accessibility, visual quality, runtime behavior, repair priority, re-check evidence |
| Motion | Choreography, ownership, fallbacks, reduced motion, performance budgets, acceptance frames |
- Platform detector:
python scripts/detect_platform.py --text "Build a WinUI settings app" - Token exporter:
python scripts/token_export.py --input assets/tokens/fluent-2.json --target winui --output Resources.xaml - Deterministic plugin archive:
npm run pack:plugin - Repository checks:
npm run check
Token export targets are css, tailwind, swiftui, compose, and winui.
SKILL.md Main router and context budget
skills/ Platform and task-specific skills
references/ Narrow decision and implementation references
commands/ Explicit Stark modes
assets/ Tokens, fonts, logo, and curated proof
scripts/ Detector, token exporter, packaging, validation
evals/ Trigger and design-quality evaluation cases
tests/ Behavioral and repository-health tests
Requires Python 3.12+ and Node.js 20+.
npm ci
npm run check
npm run pack:pluginnpm run check runs unit tests, strict marketplace lint, smoke prompts, local-link checks, context-budget checks, secret hygiene, line-ending checks, and archive validation.
See CONTRIBUTING.md, reference governance, and the review contract.
Stark is a prompt-and-reference plugin, not a component library. It works with the user's existing frontend or native stack and does not require a runtime dependency in generated applications.
The project can be distributed through npm in the future, but the current package is intentionally marked private while an install/doctor CLI and npm-specific user value are designed. The plugin archive remains the canonical artifact.