Design Engineering Fellow

Scout runs on Haiku in production. Getting a smaller, faster model to handle the full tool-use loop reliably is a prompt and tool-design problem, not a model-capability one. The agent reaches the storefront through MCP servers for what they cover well (catalog, policies, signed-in features), with custom tools filling the gaps. The catalog search tool runs server-side and overlays Storefront API data onto the MCP result before the model sees it, so the model gets richer context in one round-trip instead of two.

Product cards land inline mid-sentence at the exact reference point in the model's response. The model emits a short marker on its own line, the frontend parses it while streaming, and the card renders in place. Tap a card and it lands in the cart immediately, reconciled with the server in the background and rolled back silently on failure. Drag a card from chat into cart, and the composer, header, and cart panel all shift in proportion to the drag. What reads as one continuous gesture is three coordinated parts moving together.

The architecture is multi-store from the ground up. Adding a brand is a config file. The chat handler and components stay untouched. Everything is open on GitHub at github.com/jonathanmoore/scout: the prompt-engineering doc, architecture docs, accessibility audit, and multi-store scaffolding.

I built the Shopify theme and design system from the ground up. Performance, accessibility, and conversion-aware interaction craft were defaults from the start, not retrofits. Currently refreshing the UX for a new brand direction. Site refresh launching May 14, 2026.

Designed and shipped a Shade Finder quiz that generated $4.2M in direct revenue. The shade-matching problem hides in plain sight. A routine quiz flow, done well, removes the single biggest barrier to buying complexion products online. Quiz responses map back to the customer profile and feed downstream: retention messaging in email and SMS, retargeting in Meta, and product decisions on the merchandising side.

I built internal tools that save the CX and operations teams 20+ hours a week, using Claude Code wired into the Shopify Admin API. One of those tools is a fraud detection system that stops hundreds of fraudulent orders a month: return abuse, address cycling, reseller fraud. It's live at Kosas now and becoming a standalone product for other Shopify merchants as PrismLock.

Continuous A/B tests across acquisition and conversion surfaces, with results feeding back into the design system and the storefront. The experimentation extends past the interface. I prototyped computer vision-based shade matching using OpenCV, detecting undertones and facial regions and filtering noise like blemishes and hair to push accuracy past what a traditional quiz flow can hit. Live prototype.

Themes ship fast because the architecture is the same across them. Liquid for server rendering, Vite for builds, Web Component islands for the interactive parts, and no framework owning the page. What makes new themes faster than that architecture alone is a set of custom Claude Code skills wired into the project. Slash commands scaffold sections, blocks, and snippets with correct schema, translation across thirty languages, and accessible markup baked in, and the agent verifies accessibility in the browser as it works.

A premium Shopify theme used by 30,000+ merchants, top 1% of 1,000+ themes for Core Web Vitals. Performance and accessibility built in from the start.

Canary Yellow is the art store for the Virgil Abloh Archive. A custom Shopify build, anchored by the Air Jordan 1 OG High × V.A.A. drop. One week to ship. Design and engineering ran simultaneously while I was on a road trip across ten states. Final tweaks went through Claude Code on my phone.

A 2025 Shopify build for Yeezy. The brief came as a text message and a screen recording: "I want the website to be like this." Ambiguous direction, high expectations, no room for error. Wes Bos broke down the engineering in a 16-minute analysis after digging through the code.

An open source framework: the architectural patterns, the Claude Code skills, and the component creator that scaffolds new sections, blocks, and snippets to the same standard the rest of the practice runs on. Live at kona-theme.jonathanmoore.com.

Twice in my career, a new medium has rewritten what it means to be a designer. First in the early 2000s, then again now.

Before the web, design meant print, identity, packaging. You trained into those disciplines for years. Then digital, and especially interactive, arrived, and the work outpaced the schools. The people defining it weren't always the credentialed ones.

Eventually a clear design career path emerged in the 2010s. Follow it, get the role, progress in your career, ship clean work inside a clean process. It worked. For big teams on big roadmaps, it still does.

AI is rewriting the path again. What used to take a team and months of planning takes a week or an afternoon for one person. The designers who'll thrive now are the same ones who thrived last time: builders, with design, engineering, and product instinct in one person.

The companies that win the next decade will be the ones who hire builders, not roles. The designers who win will be the ones who can hold the idea of an entire product in their head and ship it.

On the design side, the system is the artifact. Design systems built so their patterns translate cleanly into prompts, which lets product managers, engineers, and anyone else on the team make design decisions inside the company's design language without needing me in the loop.

On the engineering side, the work runs through Claude Code: detailed plans, sub-agents executing against them, parallel terminals and git worktrees keeping multiple threads moving at once. The practice has to evolve with the tooling, so the specifics will change. What I'm figuring out now is orchestration. How long-running agents can compound the workflow rather than accelerate individual tasks.