The paper argues that product design is undergoing a fundamental transformation due to AI, shifting the focus from execution and "drawing pictures of apps" to taste, judgment, and strategic orchestration. The author's personal experience using Claude Code to generate a functional UI from their design system components in three prompts crystallized this shift, demonstrating the potential to bypass the traditional design-to-development handoff.

The author contends that the common discourse around AI and job displacement is a "wrong debate," focusing on headcount rather than the more crucial question of process. AI doesn't eliminate core functions like problem definition, design, or coding, but rather changes who performs them, their speed, and where bottlenecks occur. This is elucidated through the distinction between invisible work (e.g., coding, PRD writing, data analysis), which is easier to automate because quality gaps are hidden behind the user interface, and visible work (e.g., UI, user flows, experience), which is user-facing and where quality gaps are immediately apparent.

Using a plumbing analogy, engineering (invisible work) is likened to pipes behind a wall, where AI delivers massive velocity gains (e.g., orchestrating multiple coding agents). In contrast, software design (visible work) is "the wall" and the "tap," where intuition, user experience, and aesthetic quality are paramount. AI struggles with the latter because it excels at following standards and training data but cannot effectively integrate the nuanced, complex context derived from extensive human user research or make subjective judgments about novelty and taste.

A critical bottleneck in product development, the designer-to-developer handoff gap (translating Figma mockups into production code), is identified as "absurdly wasteful." The paper advocates for designers to "design in code, not in Figma," working directly in the final material that ships. This is becoming feasible as AI collapses the translation step. Companies like Monday.com are developing Design-System Model Context Protocols (MCPs)—machine-readable representations of components, tokens, accessibility rules, and usage patterns—and orchestrating 11-node agentic workflows to build structured context for coding agents, rather than having AI write code directly. This approach, exemplified by Anthropic's designers committing code directly to production, shifts design from parallel artifact creation to direct contribution to the shipping product.

What remains uniquely human in this AI-augmented workflow is orchestration and judgment. The bottleneck moves from AI's capability to the human's ability to direct the work, break it down, and critically evaluate outputs. Quality in AI-powered products extends beyond surface polish to encompass trust, clarity, and reliability. AI can automate the 20% of a design leader's work related to visual production and mockups, but it cannot perform the 80% dedicated to communication, alignment, justification, and strategic decision-making ("justification tax"). The most impactful application of AI is not to speed up existing tasks but to remove human limitations, enabling previously impossible capabilities. This transition disproportionately affects less experienced designers who lack the judgment to critically assess AI-generated work, thus raising the skill floor.

This paradigm shift necessitates a change in team structure. The paper suggests a move away from PM-heavy "feature factories" towards small, empowered teams (e.g., 2-3 engineers, a PM, and a designer). Design systems become critical infrastructure, serving as the backbone for AI-assisted design at scale by providing the structured context necessary for intelligent automation. The "compounding bet" implies that designers who embrace AI orchestration will operate at a fundamentally different speed and scope, shipping working UIs while others remain mired in traditional handoff processes. The future product design job emphasizes taste, judgment, and the ability to direct AI, rather than manual pixel-pushing.