ROY SHILOH
Product designer for complex AI products, clinical workflows, and technical B2B systems. Currently Head of Design at Berries, designing an AI-native clinical platform for mental-health professionals.
A brand-aware AI illustration workspace. Vinos Studio helps designers and product teams generate new illustrations that remain consistent with an existing visual language.
Product creator, design lead, frontend contributor.
I created Vinos together with a software developer, working across product strategy, UX, visual design, frontend implementation, and feature development.
AI image tools generate an image. They do not extend a brand.
Vinos started from a problem I repeatedly experienced in my own design work. AI image tools could generate an attractive illustration, but producing a full set of assets in the same style was much harder. Colors shifted, linework changed, compositions became inconsistent, and every new asset required rebuilding the prompt and uploading the same references again.
The tools were designed to generate an image, not to extend a brand. We created Vinos around a different idea:
What if a visual style could be saved and reused like any other design asset?
Style and prompt, separated.
Users upload existing illustrations, turn them into a reusable style, and use that style to generate new brand-consistent assets.
The core workflow:
- Upload visual references.
- Create and save a style.
- Describe the new illustration.
- Generate and refine the result.
- Reuse the style for future assets.
This separates two forms of intent:
- Style — how should it look?
- Prompt — what should it contain?
Instead of repeatedly describing the brand language, users can focus on the asset they need.
The create surface — the fewest possible controls in front of a model that would otherwise expose forty.
Style as a persistent object.
- Problem
- In most AI image tools, style is hidden inside a prompt or attached to a single generation. It cannot be named, revisited, or trusted to stay the same tomorrow.
- Options
- Keep style inside the prompt; attach it to a project; make it a first-class object with its own identity and lifecycle.
- Decision
- Style is a first-class object. It can be named, saved, selected, and reused across the product — the same way a color token or a component is.
- Result
- A brand's visual language becomes something the product remembers, not something the user re-describes every session.
Reference → reusable style. The unit of consistency is the style, not the prompt.
References before prompt engineering.
- Problem
- Many users can recognize the style they want more easily than they can describe it. Tools that lead with a prompt box exclude everyone who is not already a prompt engineer.
- Options
- Lead with a prompt; lead with a template gallery; lead with the user's own visual references.
- Decision
- Vinos starts with visual examples. Users show the system what they mean before they are asked to describe it.
- Result
- The product meets designers and brand teams in the medium they already work in — images, not paragraphs.
A quiet creative workspace.
- Problem
- Generative products often surround the output with dense controls, model settings, and marketing surface. The generated work stops being the focus.
- Options
- Ship every parameter as a visible control; hide advanced controls behind disclosure; design a neutral surface where the work is the subject.
- Decision
- Neutral surfaces, large previews, minimal controls. The interface stays quiet so the generated work remains the focus.
- Result
- Generation is not treated as a modal event. It is a state the interface stays in — calm, editable, iterative.
Iteration state — generation is a state the interface stays in, not a modal event.
Working across design and code, not handing off between them.
I worked directly with the developer from the initial concept through implementation. My work included:
- Defining the product concept and core workflows.
- Designing the interface and visual system.
- Prototyping interactions.
- Contributing frontend changes and pull requests.
- Building later features with Cursor and Claude Code.
- Reviewing implementation quality and refining the shipped experience.
I used Figma MCP to work more closely between the design environment and the codebase, reducing the distance between design decisions and implementation.
I also used a structured Grill Me workflow to challenge assumptions, identify weak interactions, and avoid accepting the first plausible AI-generated solution.
Early process diagrams — mapping style creation, generation flows, and interface states before building the product.
A working product, and a shift in how I work.
Vinos became a working product that connects visual references, style creation, generation, refinement, and reuse in one workflow.
The project showed me that AI products need more than a prompt box. They need persistent context that helps users understand what the system remembers and what will remain consistent.
It also changed how I work as a designer. Instead of treating code as the final destination of a handoff, I used it as another material for shaping and testing the product.