Weekly Digest // DESIGN_SYSTEMS — Week 44-2026
architectureWeekly Report

Native interaction and design evaluation | Week 44 Design Systems

Connect native select interaction with Figma’s scoped editing evaluation and UX ownership of acceptance criteria, research, and user-success signals.

calendar_todaysummarizeWeek 44-2026
TAG: ARCHITECTUREREAD_TIME: 3_MIN

A better way to handle select behaviour

Martin Underhill explains why a visually branded replacement for a native select also inherits difficult interaction decisions. In his comparison, macOS requires opening the menu and explicitly committing an option, while Windows allows closed-menu selection and treats a focused option as selected. A custom component can mix these conventions and leave users of both platforms with unfamiliar behavior. The emerging appearance: base-select model combines an open menu, a visible button, explicit selection, and non-modal interaction, giving design-system teams a clearer reference. Underhill recommends considering the customizable native control before rebuilding it; the argument concerns an interaction model, not proof that every browser implementation or custom substitute is accessible.
TAG: TOOLINGREAD_TIME: 5_MIN

How we’re training a Figma-fluent editing model

Figma describes training a model for native design edits using canvas context, conversation history, and design-system information. Its pipeline generates synthetic editing examples, filters or repairs unsuitable results, and calibrates model-based judges with designer annotations. On 445 real-world editing examples, the inspected chart reports 77.8% for Figma edits 1026 at medium effort and 76.9% for GPT 6.1 Sol at medium effort, using an automated design judge validated by humans. These are internal evaluation results without published error bars, not an independently replicated ranking across all design work. The useful method is to evaluate structural edits and visual outcomes together while keeping component and theme constraints available to the editing agent.
TAG: ARCHITECTUREREAD_TIME: 12_MIN

The Role of UX in Agentic Software Factories

Tanner Kohler argues that turning UX knowledge into agent context does not replace discovering user needs or evaluating the resulting product. His critique of AI-native development playbooks proposes user-value acceptance criteria in intent.md, shared product and UX judgment, and explicit escalation to design or research specialists. Machine-readable research and design-system assets remain valuable, but someone must own and maintain that context. After launch, task abandonment, conversion changes, and support signals should complement technical monitoring because working code can still deliver a confusing experience. This is expert guidance rather than a controlled outcome study; teams can use the proposed checks to challenge whether their approval process measures user success as well as artifact conformance.
summarizeDigest_Summary

A design system needs more than visual consistency. Martin Underhill’s select comparison shows how a familiar-looking component can still mix platform conventions and surprise users, making native interaction behavior part of the design contract.

Figma’s editing-model work turns component and theme context into inputs for generation and evaluation. Its internal benchmark offers an inspectable method, but a small score difference without error bars cannot settle broader model superiority or prove that users complete tasks successfully.

Keep structural fidelity, interaction quality, and user success as separate acceptance checks. Tanner Kohler’s UX critique extends this responsibility beyond the component library: research should shape what gets built, specialists need review paths, and post-launch monitoring should include abandonment and support signals. Machine-readable context helps agents, while ownership and evaluation remain ongoing work.

Key Takeaways
  • Include selection, focus, dismissal, and platform expectations in the design contract for form controls.
  • Read design-model scores with their task set, judge, and uncertainty before inferring broader superiority.
  • Assign UX ownership of research context, user-value acceptance criteria, specialist review, and post-launch success signals.