Design & AI Skills — Biblioteca de Referência
jpoindexter/design-and-ai-skills
40 skills de referência para design e IA, portáveis em SKILL.md para agentes Claude
Ver no GitHubColeção de 40 skills de nível production para design de produtos e engenharia de IA, organizadas em dois acervos: 27 skills de design system (tokens, tipografia, acessibilidade, componentes) e 13 skills de engenharia de IA (LLMs, agents, observabilidade). Cada skill é um arquivo SKILL.md autossuficiente com referências a padrões reais (WCAG, Material Design, RAGAS, OWASP), números concretos, tabelas de decisão e código portável para Claude Code e agentes compatíveis.
Instalação
npx convez-ai jpoindexter/design-and-ai-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
- Implementar design systems com tokens W3C e parity Figma-to-code
- Estruturar pipelines de produção para LLMs e sistemas de agents
- Revisar acessibilidade (WCAG 2.2, APCA) em interfaces de produto
- Arquitetar observabilidade, evals (RAGAS) e rastreamento de LLMs
- Documentar padrões de design escaláveis para múltiplas plataformas
Veredito

Biblioteca sólida e bem estruturada de skills técnicos com foco em padrões reais (W3C, WCAG, OWASP) em vez de dicas genéricas. Ideal para times que já têm contexto de design systems e engenharia de IA e precisam de referências confiáveis e copy-pasteable; requer esforço para integração inicial, mas entrega alto valor em manutenção de padrões.
README
A tradução para o português deste README ainda está na fila.
Design & AI Engineering Skills
40 reference-grade skills for product design and production AI engineering — self-contained SKILL.md files that load into Claude Code, Codex, or any agent that reads a skills directory.
Two libraries: 27 design-system skills covering the full UX/design canon, and 13 AI engineering skills covering production LLM and agent systems.
What these are
Each skill is a single SKILL.md file with a YAML frontmatter block and a deep reference body. The body is written as an authoritative reference, not a tips list: real numbers, formulas, comparison tables, copy-pasteable code, do/don't pairs, named failure modes, and attribution to the actual standards and papers (Nielsen/Norman/WCAG, vLLM/PagedAttention, RAGAS, OWASP LLM Top 10, W3C DTCG, and so on).
When a skill is installed, the agent loads the full content into its context at the start of the relevant task — so it works from the actual standard, not a summary of a summary.
SKILL.md format
---
name: slug-matching-folder-name
description: One-line summary used by the agent to decide relevance.
tags: [category, topic, subtopic]
---
# Title
Body — markdown, tables, code blocks.
Frontmatter fields:
| Field | Required | Notes |
|---|---|---|
name | yes | Must match the containing folder name exactly |
description | yes | Used for skill discovery; be specific |
tags | yes | Array of lowercase strings |
Body conventions used in this library:
- Sections numbered (§1, §2…) with a clear heading
- Comparison tables for any decision with >2 options
- Concrete numbers (budgets, thresholds, sizes, timings) wherever they exist
- Explicit do/don't pairs for the most common mistakes
- Named failure modes, not vague warnings
- Attribution to the originating standard, paper, or framework (with year where useful)
Repository structure
design-and-ai-skills/
├── README.md ← you are here
├── index.html ← root viewer (both libraries)
├── scripts/
│ ├── build-skills-index.py ← rebuild ai-engineering-skills/index.html
│ └── build-design-skills-index.py ← rebuild design-system-skills/index.html
├── design-system-skills/
│ ├── README.md ← skill list for this library
│ ├── index.html ← standalone viewer
│ └── <slug>/SKILL.md ← one folder per skill (27 total)
└── ai-engineering-skills/
├── README.md ← skill list for this library
├── index.html ← standalone viewer
└── <slug>/SKILL.md ← one folder per skill (13 total)
Design System Skills — 27 skills
For any screen: web, iOS/Android, macOS/Windows, tablet, TV, watch, foldable. Current to 2025–2026 standards (WCAG 2.2, Material 3, Apple HIG, W3C DTCG, OKLCH, container queries, view transitions).
Foundations
| Skill | What it covers |
|---|---|
design-tokens | W3C DTCG format, 3-tier primitive→semantic→component model, Style Dictionary v4, Figma↔code parity, multi-brand theming |
typography-system | Modular and fluid scales, clamp(), variable fonts, OpenType, baseline rhythm, accessible ramps, cross-platform tokens |
color-and-elevation | OKLCH ramps, WCAG/APCA contrast, semantic tokens, elevation-as-lightness dark mode, copy-pasteable CSS |
grid-and-spacing | 8pt spacing scale, column/baseline grids, responsive breakpoints, modern CSS Grid and container queries |
iconography-and-imagery | Icon grids, optical sizing, variable axes, SVG delivery, aspect ratios, art direction, modern image formats |
Structure
| Skill | What it covers |
|---|---|
atomic-design | Brad Frost's atoms→molecules→organisms→templates→pages, Figma/code component tree organization, abstraction levels |
layout-and-composition | Visual hierarchy, Gestalt, alignment, whitespace, scanning patterns, intrinsic CSS (Grid/Flexbox/subgrid/:has) |
components-and-states | Component anatomy, full interaction-state matrix, variants vs sizes vs states, sizing tokens, headless/data-state patterns |
design-system-governance | Operating models, contribution and release process, Figma↔code parity, documentation, adoption metrics, anti-patterns |
design-system-frameworks | Material 3, Apple HIG, Fluent 2, Carbon, Polaris, Primer, Ant, Spectrum; Tailwind, shadcn/ui, Radix, React Aria; adopt-vs-adapt decision |
Interaction & device
| Skill | What it covers |
|---|---|
interaction-and-motion | Duration scales, easing tokens, choreography, view transitions, scroll-driven animations, @starting-style, prefers-reduced-motion |
forms-and-data-entry | Single-column layout, persistent labels, input type + keyboard + autocomplete per field, :user-invalid validation, multi-step wizards |
responsive-and-multi-device | Mobile-first breakpoints, fluid clamp() type, container queries, logical properties, dvh/safe-area, phone/tablet/TV/watch/foldable |
platform-conventions | iOS/HIG, Android/Material 3, macOS, Windows 11 Fluent — concrete numbers, comparison tables, adapt-vs-custom-brand decision |
UX & inclusion
| Skill | What it covers |
|---|---|
usability-heuristics | Nielsen's 10 heuristics, Norman's fundamentals, Shneiderman's 8 golden rules, Rams' 10 principles, heuristic evaluation method |
laws-of-ux-and-psychology | Fitts, Hick-Hyman, Miller, Jakob, Tesler, Postel, Doherty, Peak-End, behavioral econ cluster — each with origin, UI application, design rule, pitfall |
ux-writing-and-content | Voice and tone systems, every UI string type with before/after, terminology consistency, inclusive language, localization readiness |
accessibility-and-inclusive-design | WCAG 2.2 AA/AAA, APCA, keyboard/screen-reader/touch/motor, forms, motion/cognition, testing, EAA/Section 508 legal floor |
inclusive-design | Microsoft's 3 principles + Persona Spectrum, social model of disability, cognitive/neurodivergent/sensory/cultural/economic/age patterns beyond WCAG |
information-architecture-and-navigation | Organization/labeling/navigation/search systems, taxonomy depth-vs-breadth, nav patterns per device, wayfinding, card sort/tree test/first-click |
data-visualization | Perceptual encoding ranking, chart-by-intent selection, data color scales, Tufte's data-ink discipline, accessible charts, interaction patterns |
ethical-design-and-dark-patterns | Brignull/Mathur dark-pattern taxonomy, every pattern with example + fix, consent/privacy UX, attention ethics, FTC/GDPR/DSA legal exposure |
performance-and-perceived-speed | Core Web Vitals (LCP/INP/CLS budgets and fixes), optimistic UI, skeletons, prefetch-on-intent, loading states, field vs lab measurement at p75 |
onboarding-and-empty-states | Every view state (empty/loading/error/offline/success/blocked), first-run flows, show-don't-tell, seed data, activation to aha moment |
multimodal-voice-and-haptics | Touch/visual/voice/audio/haptic modality combinations, conversational/LLM-agent UX, voice error recovery, iOS/Android haptics, TV/watch/AR |
internationalization-and-localization | Text expansion, RTL/bidi with CSS logical properties, Intl formatting, ICU pluralization, script typography, pseudolocalization, translation workflow |
brand-identity-and-design-principles | Opinionated design principles, logo systems, attributes-to-visuals translation, brand-in-product, brand→token bridge, guidelines, rebrands |
AI Engineering Skills — 13 skills
For engineers shipping LLM and agent systems in production. Grounded in specific systems and papers: vLLM/PagedAttention, FlashAttention, AWQ/GPTQ/SmoothQuant, RAGAS, OWASP LLM Top 10, OpenTelemetry GenAI.
Context & harness
| Skill | What it covers |
|---|---|
harness-and-context-engineering | The control loop, tools, memory, retries, budgets; context as a managed token budget; lost-in-the-middle, context rot, just-in-time retrieval, compaction, sub-agent isolation |
adaptation-strategies | Choosing between ICL, RAG, fine-tuning (LoRA/QLoRA/DPO), and distillation; data requirements; catastrophic forgetting; combining methods |
Inference & serving
| Skill | What it covers |
|---|---|
inference-performance | TTFT/TPOT/throughput, prefill vs decode, the roofline model, continuous batching, PagedAttention, chunked prefill, disaggregation, FlashAttention, GPU sizing |
inference-caching-and-kv | Anthropic cache_control, OpenAI prefix caching, Gemini implicit/explicit; semantic caching; KV-cache internals (PagedAttention, RadixAttention, eviction, multi-tenant safety) |
quantization-and-model-compression | FP8/INT8/INT4, GPTQ/AWQ/SmoothQuant/NF4/GGUF k-quants, KV quantization, speculative decoding (Medusa/EAGLE/n-gram), distillation, quality cliffs |
Reliability & agents
| Skill | What it covers |
|---|---|
structured-output-and-tool-calling | JSON mode, constrained/grammar-guided decoding, json_schema coercion, Zod/Pydantic validation, validate→repair→escalate→fallback loop, idempotency for retries |
agent-reliability-and-guardrails | Plan→act→observe failure physics, five budgets (iteration/tool-call/token/wall-clock/cost), runaway prevention, circuit breakers, three guardrail layers, checkpoint/resume |
model-routing-and-fallback | Routing by cost/latency/quality, cascade escalation, multi-provider fallback, circuit breakers, hedging, degraded-mode UX |
Retrieval & evals
| Skill | What it covers |
|---|---|
rag-architecture | Full ingest→chunk→embed→index→retrieve→rerank→assemble→generate pipeline, chunking strategies, hybrid search, reranking, context assembly, failure modes |
llm-evals-and-retrieval-quality | Golden/regression/adversarial eval sets, LLM-as-judge biases, recall@k/MRR/nDCG, grounding/faithfulness/attribution, RAGAS scoring, CI gates |
Ops, cost & safety
| Skill | What it covers |
|---|---|
llm-observability-and-cost | Distributed traces and spans, OpenTelemetry GenAI conventions, drift detection, per-feature/tenant/user cost attribution, LangSmith/Langfuse/Phoenix/Helicone |
llm-safety-and-multitenancy | OWASP LLM Top 10, direct/indirect prompt injection, privilege separation, dual-LLM pattern, data-leakage prevention, cross-tenant cache/index/memory isolation |
production-failure-modes-and-tradeoffs | Full failure taxonomy (symptom/root-cause/detection/mitigation), latency/quality/cost/reliability tradeoff map, production-readiness checklist, incident response |
Install
Fresh clone (recommended)
git clone https://github.com/jpoindexter/design-and-ai-skills
cd design-and-ai-skills
./bootstrap.sh
bootstrap.sh does two things in one step:
- Installs all 40 skills to
~/.claude/skills/ - Wires a
post-mergegit hook so futuregit pulls auto-reinstall
Reload Claude Code after running. You're done — no manual steps needed after this.
Existing clone / manual install
./install.sh
Finds every SKILL.md under design-system-skills/ and ai-engineering-skills/ and copies each to ~/.claude/skills/<slug>/. Use this if you cloned before bootstrap.sh existed, or want to reinstall without pulling.
Preview what will be installed first:
./install.sh --dry-run
Keeping skills up to date
Once bootstrap.sh has been run once, git pull handles everything:
git pull # post-merge hook fires automatically → all 40 skills reinstalled
No manual install.sh needed after a pull.
Install a single skill
cp -R design-system-skills/usability-heuristics ~/.claude/skills/
Other agents
Each <slug>/ folder contains one SKILL.md. Copy the folder into whatever directory your agent reads for skills. The format (frontmatter + markdown body) is not Claude-specific — any agent that can read a skills directory with YAML frontmatter will work.
Browse
Each library ships a self-contained index.html viewer — no server, no build, works from file://.
| Viewer | Path |
|---|---|
| Both libraries | docs/index.html |
| Design system skills | docs/design-system-skills.html |
| AI engineering skills | docs/ai-engineering-skills.html |
Open any of them directly in a browser. Grouped sidebar, filter box, collapsible skill cards.
Regenerate the viewers
After adding or editing a skill, rebuild the corresponding index.html. Requires mistune:
pip install mistune
# Rebuild AI engineering viewer → docs/ai-engineering-skills.html
python3 scripts/build-skills-index.py ai-engineering-skills
# Rebuild design system viewer → docs/design-system-skills.html
python3 scripts/build-design-skills-index.py
Adding a skill
- Create
<library>/<slug>/SKILL.md—slugmust be kebab-case and match thename:in frontmatter. - Fill in the frontmatter (
name,description,tags). - Write the body. Conventions: numbered sections, comparison tables for multi-option decisions, concrete numbers where they exist, attribution to originating standards/papers, named failure modes, do/don't pairs.
- Rebuild the viewer for that library (see above).
- Add the skill to this README and to
<library>/README.md.
License
MIT — use, adapt, share.