Design System Skills
delayed-byte/design-system-skills
Skills para IA construir UIs coerentes com design system como fonte de verdade
Ver no GitHubConjunto de skills portáteis que permitem ao Claude, Codex e Cursor construir interfaces coerentes usando um design system como fonte única de verdade. Inclui prompts reutilizáveis para definir briefs, gerar telas, implementar código e enforçar consistência visual em todo o projeto.
Instalação
npx convez-ai delayed-byte/design-system-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar múltiplas telas mantendo coerência visual e estilo
- Implementar componentes UI respeitando tokens de design
- Documentar e enforçar padrões de design em projetos colaborativos
- Reduzir tempo de restyling e iteração de UI com IA
Veredito

Solução prática e bem estruturada que resolve um problema real: IA tendendo a reinventar design em cada geração. O framework de quatro skills (brief → generate → build → enforce) é maduro e documentado, com validação em CI. Recomendado para times que usam IA para frontend e precisam de consistência visual.
README
A tradução para o português deste README ainda está na fila.
design-system-skills
A small collection of portable AI-agent skills (Claude Code, Codex, Cursor, and any tool that reads the open SKILL.md format) that make AI build a coherent frontend by treating a design system as an input — a source of truth the model references instead of re-inventing pixel-level decisions on every screen.
The skills, and the workflow they encode, are adapted from the X article "4 prompts that make Claude Code build your best UI" by @Mnilax.
The problem these skills solve
When you ask an AI to "build a settings page," it draws a working page. Ask for the next screen and it draws that one from scratch — a new blue, a new card radius, a new idea about spacing — because it has no memory of the decisions it made the first time. Across ten screens you don't get a product; you get ten drawings that happen to share a folder.
The fix isn't a better model, it's a source of truth: one place that says this is the primary color, this is the button and its states, this is the spacing scale, and here is when to use a drawer vs a modal. Hand the model that and it stops inventing and starts referencing. The reported effect of doing this: on-brand-on-first-generation screens went from 2/10 to 9/10, restyle time from ~6 hours to ~40 minutes, and AI-invented one-off hex values from 19 to 0.
The four skills
Each skill is a self-contained folder with a SKILL.md (the why, a reusable prompt template, a worked example, and the mistake it prevents). They map to the four stages of the pipeline:
| Skill | Stage | What it does |
|---|---|---|
| design-system-brief | 1. Define | Write a specific, build-ready brief so the generated design system isn't generic. Specificity early is the whole game. |
| generate-screens-with-design-system | 2. Generate | Generate a full set of screens in one pass with the system attached, so nav/cards/charts stay consistent across all of them. |
| build-ui-from-design-system | 3. Build | Build a screen/component in code using only the system's tokens — and "stop and ask" rather than invent when something isn't covered. |
| enforce-design-system | 4. Enforce | Set a standing project rule (CLAUDE.md / Cursor rule) so every future generation references the system and the build never drifts. |
They're designed to be used in sequence, but each works on its own.
Repository layout
design-system-skills/
├── README.md # this file
├── CONTRIBUTING.md # collaborator contribution guide
├── CODE_OF_CONDUCT.md
├── package_skills.sh # rebuild the *.skill zips from sources
├── tools/
│ └── validate_skills.py # frontmatter checks (run in CI)
├── .github/ # CI workflow + issue/PR templates
└── skills/
├── design-system-brief/
│ └── SKILL.md
├── generate-screens-with-design-system/
│ └── SKILL.md
├── build-ui-from-design-system/
│ └── SKILL.md
├── enforce-design-system/
│ └── SKILL.md
├── *.skill # zipped, installable versions of each skill
└── evals/ # triggering tests (see below)
├── routing-eval.json # the eval query set + intended routing
├── judge_results.json # raw blind judge outputs (3 rounds)
├── score_triggering.py # scores judge results vs intended routing
└── triggering-report.md # written findings
Each folder is the editable source. Each matching *.skill file is a zip of that folder — the installable artifact (use the "Save skill" button in the Claude desktop app, or install via Settings → Capabilities).
Installing & using
These skills use the open SKILL.md standard (name + description + body, with progressive disclosure), so the same files work unmodified across Claude Code, Codex, Cursor, and other agents — none of the four are tool-specific.
- Install a skill by copying its folder into the agent's skills directory:
- Claude Code / Claude desktop → use the "Save skill" button on the
.skillfile, or Settings → Capabilities. - Codex → drop the folder in
~/.agents/skills/(available everywhere) or.agents/skills/(committed to a repo, shared with the team). - Cursor / others → their skills directory; the format is identical.
- Claude Code / Claude desktop → use the "Save skill" button on the
- The skill triggers automatically when your request matches its description — e.g. "help me set up a design system for a new app" pulls in
design-system-brief. - For step 4,
enforce-design-systemproduces a standing rule you drop into your agent's persistent-instructions file —CLAUDE.md(Claude Code),AGENTS.md(Codex), or a Cursor project rule — so consistency persists across every future build without re-stating it.
The pipeline in one glance
brief ──► generate screens ──► download/export system ──► build in code ──► standing rule
(define) (generate) (hand off) (build) (enforce)
Evals — how triggering was tested
Because these four skills overlap heavily, the real risk isn't whether each fires but whether requests route to the right one. The skills/evals/ folder documents that test.
- Method. The skill-creator's automated harness (
run_eval.py) shells out to a logged-inclaudeCLI, which wasn't available in the build environment. Instead, each query was scored by 3 independent fresh-Claude "judges" shown all four descriptions, asked which one (or NONE) each query triggers. This directly measures routing between the skills. - Results. 41/42 correct overall (98%). Realistic queries: 24/24. Hard boundary cases: 9/10 (the one miss was a CI/lint "pre-commit hook" request over-triggering
enforce-design-system). That description was then tightened to exclude CI/lint/git-hook tooling, and the re-test passed 8/8. - Reproduce it.
python3 skills/evals/score_triggering.pyreadsjudge_results.jsonand prints accuracy, unanimity, trigger rate, and any misroutes. For ground-truth production numbers, re-run the skill-creator'srun_eval.pyagainstrouting-eval.jsonin an environment with an authenticatedclaudeCLI.
The descriptions discriminate cleanly on the three axes that actually separate these skills:
- Does a system already exist yet? →
design-system-brief(no) vsenforce-design-system(yes) - Design tool or code? →
generate-screens-with-design-systemvsbuild-ui-from-design-system - One-time build or standing rule? →
build-ui-from-design-systemvsenforce-design-system
Editing a skill
Edit the SKILL.md in the skill's folder, then rebuild the installable zips:
./package_skills.sh # rebuilds every skills/*.skill from its source folder
The description field in the frontmatter is what controls triggering — if a skill fires too often or not enough, that's the line to tune. After editing, validate and re-check routing:
python3 tools/validate_skills.py
python3 skills/evals/score_triggering.py
Contributing
This repository is not accepting unsolicited contributions. Authorized collaborators should follow CONTRIBUTING.md and the Code of Conduct.
Credits
Workflow and prompts adapted from "4 prompts that make Claude Code build your best UI" by @Mnilax. The design-system tool referenced throughout is Moonchild; the skills themselves are tool-agnostic and work with any system that can export tokens/components/guidelines.
Copyright
Copyright © 2026 @delayed-byte. All rights reserved.
This repository is provided for review and portfolio purposes only. No permission is granted to copy, modify, distribute, or use this code without written permission.