The New Full-Stack Designer

AI tools have fundamentally shifted the design process from hand-crafted execution in Figma to AI-driven iteration, turning designers into curators/editors w

Abstract

This synthesis examines a practitioner account of how generative AI coding agents have restructured professional interface design practice. Drawing on the experience of a design studio founder now working within Shopify's Product Design Studio, it documents a migration of primary design labour from canvas-based tools such as Figma toward terminal-resident agents such as Codex and Claude Code. The central finding is that execution capacity has ceased to be the binding constraint on design output; selectivity, decomposition, and taste have replaced it. Evidence is drawn from two instrumented prototyping case studies - a procedural grass shader requiring 91 prompt iterations across roughly 217,000 rendered variations, and a physics-driven carousel tuned via live parameter exposure - alongside a set of applied interface critiques. Practical implications include the primacy of prototypes over written specifications as organisational artefacts and the emergence of "full-stack" role convergence across design, engineering, and product management.

1. Introduction

The introduction of competent code-generating agents into the design toolchain raises a question distinct from the familiar one of automation displacing labour: when execution becomes inexpensive, what remains of the discipline? The account synthesised here advances a specific answer. Agentic design - the practice of specifying, critiquing, and iterating on machine-generated interface code rather than authoring layouts directly - does not eliminate the designer's role but relocates it to curation, problem decomposition, and the setting of aesthetic standards.

Three terms require definition. Agentic iteration denotes a loop in which a human supplies references, critiques output, and re-prompts, with the artefact itself (running code) serving as the medium of communication. Parameter exposure refers to the practice of surfacing a prototype's internal variables as manipulable controls - exemplified by the tool DialKit - enabling empirical tuning in place of further textual instruction. Progressive disclosure describes the structuring of information so that depth is revealed on demand rather than presented simultaneously, as operationalised in the Primal content system.

The central claim examined here is that AI tooling has raised "the ceiling" of what is executable while leaving unchanged, or even sharpening, the designer's responsibility to decide what is worth building. The analysis proceeds as follows. Section 2 establishes the practice context from which the observations derive. Section 3 presents core findings across workflow, organisational, and craft dimensions. Section 4 isolates reusable technical insights. Section 5 situates these within broader industry trends, and Section 6 states practical takeaways.

2. Background and Related Work

The observations originate from two institutional settings. The first is a Brooklyn-based independent studio, co-founded with Jay Tal and operated for approximately two years, serving clients including Collins, Apple, and Google while deliberately limiting engagement to two or three concurrent clients to preserve focus. Within that practice, the Primal system organised website information into sequential "chapters" on the homepage mirroring top-level navigation items - an operationalisation of progressive disclosure at the level of site structure.

That studio's methodology also formalised the show-stopper component: one or two interactive elements built to a markedly higher standard of polish than their surroundings, intended to arrest attention and, in an agency context, to demonstrate craft capability to prospective clients. A representative instance is a bookshelf component on the Collins website featuring hand-tuned bounce physics implemented by a dedicated engineer (Jesper), accompanied by ambient background gradients computationally matched to on-screen colours. The second setting is Shopify's Product Design Studio, where roughly a year has been spent redesigning the company's most consequential surfaces: the Sidekick conversational AI layer, merchant admin dashboards, and the mobile application, under a stated organisational mandate ("we're entering a new era and that future needs to be designed") that elevated design's institutional standing.

3. Core Analysis

3.1 The Relocation of Design Labour from Canvas to Terminal

The primary workflow shift documented is a move away from Figma as the locus of final-fidelity design toward its use as a lightweight sketching surface, with Codex and Claude Code assuming responsibility for producing polished, functional output. This has a direct consequence for professional identity: practitioners describe themselves as "editors" selecting among AI-generated candidates rather than authors constructing each element by hand. Because agent output is no longer scarce, the binding constraint shifts to judgment - deciding which outputs merit refinement - and to the decomposition of complex interface problems into subtasks an agent can execute iteratively.

3.2 Organisational Consequences at Shopify

At Shopify, this shift manifests institutionally through two parallel initiatives: a conventional admin redesign conducted in Figma and web code, and the Sidekick team's effort to let merchants manipulate interfaces conversationally rather than through direct manipulation. Notably, prototypes built directly in a development UI playground are reported to have propagated through the organisation faster than written product requirement documents (PRDs), suggesting that executable artefacts are displacing specification documents as the primary unit of internal communication when the cost of producing a working prototype approaches the cost of describing one in prose.

3.3 Instrumented Prototyping: The Grass Shader Case

A detailed case study - recreating real grass as a web shader - illustrates both the capability and the limitations of agentic iteration. An initial attempt fed a video of grass into Codex directly; failures were diagnosed using ChatGPT as a secondary, independent agent, revealing that the model required an explicit topographical depth map of the underlying terrain before it could render grass blades convincingly. This indicates that agentic image-to-code translation does not spontaneously infer latent structural information (terrain depth, light source, shadow direction) from a reference video; such structure must be decomposed and supplied explicitly. The resulting prototype rendered approximately 217,000 distinct grass-blade variations differing by increments as small as 0.7-0.8 in a single parameter, reached after 91 prompt iterations, after which DialKit was used to expose underlying variables (length, wind, "pull/boost") as sliders for direct empirical tuning.

3.4 Applied Critique as a Durable Design Competency

A series of website reviews demonstrates that classical design heuristics remain the evaluative framework even as execution is automated. Recurring diagnoses include competing calls-to-action without a clear focal point (described via the German design-school phrase that a poster "doesn't dictate"), information density that should be condensed into a "three-up" of top features with tappable depth, and premature gating of value behind sign-up walls. Positively evaluated examples (e.g., Mentlify) were credited with restraint, generous whitespace, and scroll-depth-prioritised information hierarchy - properties attributable to human curatorial judgment rather than generative capability.

4. Technical Insights

Several implementation-level findings generalise beyond the specific case studies. First, agentic shader generation benefits from explicit decomposition of a visual reference into component signals - colour/hex values, light source direction, shadow geometry, and depth maps - rather than a single holistic prompt; without this, iterative refinement tends to stall. Second, DialKit, installed via npm run and integrated into an existing project, exposes internal state as live-adjustable sliders, which surfaces structural errors in a prototype that purely textual, chat-based iteration cannot reveal; this was applied both to the grass shader and to tuning drag/collision ("pull boost") physics in a carousel inspired by Mike Matas's portfolio work. Third, the open-source Blossom Carousel component, supporting swipe, scroll, and trackpad interaction with stacked mobile layouts, is cited as a reusable building block, indicating that even in an agent-driven workflow, hand-built open-source primitives retain value as scaffolding. Fourth, scroll-jacking (hijacking native scroll behaviour for custom animation) is identified as a persistent anti-pattern across the industry, independent of how the interaction was built, because it violates decades of platform-conditioned user expectations for interruptible, native scroll responsiveness.

5. Discussion

The findings suggest that AI coding agents function less as autonomous designers than as high-throughput execution partners whose principal advantage is tolerance for iteration volume inaccessible with human collaborators - the explicit comparison being that an engineer would "flip out" after 91 requested revisions where an agent will not. This reframes the designer's comparative advantage away from manual craft and toward the capacity to recognise quality, specify reference material precisely, and decompose ambiguous aesthetic goals into testable subtasks. The repeated emphasis that agents "execute complexity" but possess "no taste" positions human judgment as the remaining non-automatable function, consistent with broader claims across creative-tool domains that generative systems expand the feasible solution space without altering the criteria by which solutions are selected.

A notable organisational implication is the apparent displacement of PRDs by working prototypes as the fastest-propagating artefact inside a large company, which raises unresolved questions about documentation, versioning, and cross-functional alignment when executable code rather than prose becomes the primary specification medium. This account does not address how such prototype-centric communication scales to teams without agent-fluent practitioners, nor how design critique practices adapt when every stakeholder can generate plausible-looking alternatives.

6. Conclusion

This synthesis documents a shift in which the designer's role contracts around judgment, decomposition, and taste as AI agents absorb the execution burden previously requiring either personal technical skill or dedicated engineering collaborators. The grass-shader and carousel case studies demonstrate that effective agentic iteration depends on explicit structural decomposition of references and on tools like DialKit that make internal parameters directly manipulable, rather than on prompting alone. Practically, teams adopting agentic design workflows should prioritise building reusable debugging infrastructure for parameter exposure, treat working prototypes as primary specification artefacts, and invest design leadership attention in critique and prioritisation rather than manual execution.

Sources

• The New Full-Stack Designer - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub