The handoff between design and engineering has long been the most fragile link in the product development chain. For years, the industry standard has followed a rigid sequence: a designer spends days refining a high-fidelity mockup in Figma, documents every interaction in a spec sheet, and then presents it to a developer who must determine if the vision is technically feasible. This process is often a game of telephone, where the original intent is lost in translation, and the final product is a compromise between a static image and a technical constraint.

The Shift to Code-First Prototyping

At Jane Street, a designer named Edwin is dismantling this traditional pipeline by replacing static mockups with functional, code-based prototypes powered by Claude. Rather than drawing a button in Figma and writing a document to explain how it should behave, Edwin now uses Claude to build the actual feature within the codebase. This shift was most evident during a recent project to integrate LLM prompting capabilities into JSQL, Jane Street's internal SQL dialect.

Instead of iterating on a visual canvas, Edwin used Claude to refine the granular details of the submission button, implement keyboard shortcuts, and calibrate the phrasing of confirmation messages in real-time. The overhead of maintaining a Figma component library or formatting specification documents has been entirely eliminated. What began as a tool for minor UX tweaks has rapidly scaled in scope. Over the last two months, the complexity of these AI-assisted contributions has grown to include data model and library changes involving diffs of more than 2,000 lines of code. For some new applications, the Figma stage has been bypassed entirely, moving directly from a conceptual prompt to an interactive prototype.

From Persuasion to Experience

This transition represents a fundamental pivot in how product decisions are made, moving the center of gravity from persuasion to experience. In the legacy workflow, a designer's primary goal was to convince a developer that an idea was worth the effort of implementation. This created a tension where developers were hesitant to spend time building a prototype for an unproven concept, and designers struggled to prove the value of a nuance that a static image could not convey. By delivering a working Proof of Concept, the designer removes the need for persuasion. Colleagues no longer read a document to understand a feature; they use the feature to evaluate it.

This shift is made possible by the collapse of the technical barrier. While Edwin possessed a background in React, the specialized environment at Jane Street—utilizing OCaml and the internal Bonsai framework—presented a steep learning curve that would typically keep a designer out of the codebase. Claude acts as a translator, allowing a designer to experiment directly within the medium of the final product without needing to master the intricacies of a functional programming language.

However, this efficiency introduces a new cognitive risk: the iterative trap. When a designer relies on AI to generate iterations, there is a danger of staying within the boundaries of what the model suggests. This can lead to a local maximum where the design is polished and functional but lacks the creative leap that comes from thinking outside the constraints of an existing codebase. While AI is an exceptional tool for refining a mature product, it may inadvertently constrain the raw, divergent thinking required for truly novel creation.

Furthermore, the delivery of a fully baked prototype creates a unique review risk. When a reviewer is presented with a feature that already works perfectly, they are more likely to treat it as a finished product rather than a proposal. This risks turning the design review into a mere code review, where the reviewer focuses on syntax rather than questioning whether the feature should exist in the first place. It mirrors a common failure in design agencies where a client asks a designer to make a wireframe look pretty rather than questioning the underlying logic.

To counter this, the team at Jane Street has redefined the nature of the prototype. They treat these AI-generated features as disposable living documents. The code is explicitly labeled as temporary and intended to be thrown away. This framing forces the reviewer to focus on the user experience and the conceptual design rather than the implementation. Once the UX is validated, a professional engineer takes the lead to reimplement the feature into production-grade code, ensuring that quality, ownership, and architectural integrity are maintained.

The emergence of the working prototype as a communication tool suggests a future where the boundary between design and engineering is not just blurred, but irrelevant. By reducing the cost of the handoff to near zero, organizations can move from debating a vision to experiencing it, provided they maintain the discipline to treat AI output as a starting point rather than a final destination.