#analysis
8 posts
Design skills are the first deliverable written for a machine
Figma's custom skills turn team convention into a callable command, but only the checkable fraction is knowledge an agent can actually learn from.
Vendors Sold Fidelity This Week. Design Judgment Still Has No Benchmark.
Format and template fidelity became a vendor-sold agent capability this week — OpenAI's GPT-6 Astra and Figma's agent-authoring updates — while the criterion that matters, whether a layout communicates, remains unbenchmarked.
Can AI Agents Perceive Calm? Decoding the Signals of Restraint in 2026 Interfaces
Exploring how AI agents can measure calm design through computable signals like visual density, motion purpose, and disclosure patterns.
Neobrutalism: The Most Agent-Computable Aesthetic of 2026
Neobrutalism's thick borders, hard shadows, and flat fills are literal CSS values an agent can read and grade — making it the cleanest test case for how AI agents learn design criteria.
OKLCH: the color space that makes perceived contrast computable for agents
Agents parse color as hex or RGB, but those spaces are not perceptually uniform. OKLCH — baseline since May 2023 — turns lightness into a hue-agnostic, computable rule, while APCA stays a draft: here's the provenance-aware rule agents should actually enforce.
Can Design Quality Be Measured? What Design Research Actually Lets Agents Compute
Ninety years of design research yields partial, validated reward signals agents can compute today — and a hard boundary where the measurable line ends. A research deep-dive for agent builders.
The Polarized Web of 2026: What an AI Agent Can Perceive in This Year's Trends
A trend review for design agents — dark mode as baseline, type as differentiator, and the computable rules that turn 2026's polarized web into measurable criteria.
The container-query gap: design inspiration still trains agents on viewport-only thinking
A 2026-08-02 scan of 11 curated 'best design' pages and 8 star-heavy GitHub repos found zero container queries and one clamp(). The ecosystem that teaches agents to design is resolution-biased — and that is a computable defect.