(BUSAN.SASHIMI)
Weekly Digest // DESIGN_SYSTEMS — Week 35-2026
architectureWeekly Report

Alt Text Quality Escapes the Checkbox — Week 35 Design Systems

GitHub, Deque, and Allegro show design systems separating deterministic checks from contextual judgment while preparing components for AI consumers.

calendar_todaysummarizeWeek 35-2026
TAG: ACCESSIBILITYREAD_TIME: 7_MIN

Your Alt Text Passes Automated Checks. That Doesn't Mean It's Good

GitHub separates alt-text facts from judgments. Five deterministic rules flag empty-quality strings, filenames, placeholders, and nearby repetitions without credentials or network calls; an optional vision-model rule receives the image, nearest heading, title, caption, link context, and up to 600 characters of prose. WebAIM's 2026 scan found 16.2% of images missing alt text and another 10.8% with vague or duplicated values. The design favors conservative defaults because a noisy accessibility rule will simply be disabled.

Read Articlearrow_forward
DESIGN WORKFLOW53:55

6 Top Designers Show How They Actually Use Cursor

Six designers show Cursor workflows that progress from contained Figma-plugin exercises to prototypes running against real product code. At Descript, using the local production codebase means interactions inherit the actual design system, components, and edge cases instead of approximating them in a disconnected mockup. Natural-language edits help explore timeline behavior and repair small UI bugs, but the designers still review diffs and make prototypes production-ready. The transferable pattern is to begin in a bounded sandbox, keep the branch current, and use shared code as a collaboration artifact rather than a disposable demo.

AI_INFOGRAPHIC
6 Top Designers Show How They Actually Use Cursor — infographic
WATCH_VIDEOarrow_outward
AI DESIGN WORKFLOW18:39

New way to prompt Fable 5 to design amazing websites

Flux Academy follows a designer with no traditional coding background through a six-step process for a functional e-commerce prototype. The “perfect prompt” begins with a design brief: audience, brand position, growth model, technology constraints, and the choice of commerce stack are researched before generation. Claude then produces a structured project folder and technical build prompt, while the designer supplies logos, palette, type, product references, and image direction. The lesson is that Fable 5’s one-shot result depends on context architecture and curated assets, not a magic sentence.

AI_INFOGRAPHIC
New way to prompt Fable 5 to design amazing websites — infographic
WATCH_VIDEOarrow_outward
UI ENGINEERING64:16

Stop Building AI Slop – Build High-End Web Apps with AI

This freeCodeCamp course treats generic AI interfaces as a context and iteration failure rather than an unavoidable visual style. Its five-part workflow gathers a durable product specification, studies a proven interface architecture, rebuilds it for a distinct use case, applies explicit restyling tools, and adds scroll-driven polish before deploying to Cloudflare. Decisions are logged so a long agent conversation does not silently lose requirements. The useful boundary is structure before decoration: borrow interaction patterns, but replace brand, hierarchy, content, and visual language so the result is coherent rather than a clone.

AI_INFOGRAPHIC
Stop Building AI Slop – Build High-End Web Apps with AI — infographic
WATCH_VIDEOarrow_outward
TAG: COMPONENT ACCESSIBILITYREAD_TIME: 4_MIN

Axe-core Tests ElementInternals Semantics

Axe-core 4.13.0 recognizes ElementInternals for aria-allowed-attr, aria-prohibited-attr, required-parent and required-children checks, plus list and listitem rules. That prevents a correctly self-describing custom element from being reported as if its role were missing. A single browser-native component can serve React, Vue, and Angular consumers, reducing the accessibility drift of parallel implementations. Developers must explicitly grant Axe-core access to ElementInternals and continue manual interaction testing.

TAG: AI GOVERNANCEREAD_TIME: 4_MIN

How Allegro Builds AI-First Design Systems

Allegro treats language models as first-class design-system users. Its approach encodes component names, structural constraints, accessibility rules, and examples so generated interfaces stay consistent, then replaces repeated ad-hoc prompts with internal tools and plugins. Designers move toward written intent, product strategy, and business trade-offs. The counter-risk is organizational: one-person AI delivery can move quickly while disconnecting from shared review and fragmenting the product.

summarizeDigest_Summary

GitHub's alt-text scanner starts from a hard boundary: presence can be proven, quality is contextual. WebAIM found 16.2% of images on the top million home pages missing alt text and another 10.8% using vague, filename-like, or duplicated text. GitHub therefore runs five deterministic rules by default and keeps a vision-model judgment opt-in, supplied with layout and up to 600 characters of surrounding context.

Axe-core 4.13.0 now recognizes ElementInternals semantics for the rules most likely to create false positives in custom elements. That gives cross-framework component libraries a testable accessibility layer inside the element rather than requiring every React, Vue, or Angular consumer to bolt ARIA on afterward. Access to ElementInternals must still be enabled explicitly, and automated results remain only one layer of validation.

Allegro extends the consumer model again by treating language models as first-class users of its design system. Canonical names, constraints, accessibility rules, and internal tools make probabilistic generation more predictable than ad-hoc prompts. The organizational risk is a fast one-person delivery silo, so shared review and product strategy remain part of the system.

The selected videos show the same contracts at generation time. Designers get higher-fidelity prototypes when agents operate inside the real component system; strong Fable and Claude results begin with a researched brief and structured asset folder; and polished AI-built interfaces keep decisions in a durable spec before restyling. A model-ready design system is therefore context infrastructure, not merely a component gallery.

Key Takeaways
  • Keep deterministic accessibility rules cheap and enabled by default; make model-based quality judgments opt-in, context-rich, redacted, and calibrated against an offline evaluation set.
  • Upgrade to Axe-core 4.13.0 or later when testing ElementInternals, explicitly expose the internals to the scanner, and retain keyboard and assistive-technology review.
  • Document component names, constraints, and accessibility behavior in model-readable form, then wrap common generation paths in shared tools and human review.