AI image generation only stays consistent when the person directing it thinks like a designer. Conserva is what that looks like: a structured prompting framework producing a coherent visual library across an entire premium product, paired with a system-first design language built to hold its point of view as it scales.
Explore interactive prototype
Food platforms tend to fall into two patterns. Some are purely functional: structured, efficient, and difficult to browse with any sense of pleasure. Others lean heavily into visual richness, but lose clarity under the weight of it.
Neither serves the discerning home cook. What that user wants is a platform that feels like it has an opinion, on what's worth eating, how food should be described, what's worth showing and what's worth leaving out. And critically, that opinion has to hold. A platform that's expressive on the discovery screen and transactional three taps in doesn't feel curated. It feels inconsistent. In a premium context, inconsistency reads as carelessness and a user who stops trusting the curation stops using the platform.
Design challenge: How do you build a product that feels expressive and opinionated at every touchpoint, without that expression collapsing into inconsistency as it grows?
Early exploration made one thing clear: a distinct visual voice only stays distinct if it's anchored by rules: colour, typography, composition, and how those elements relate to each other across screens. Without that, the system fragments as content is added.
The instinct
Build the brand first, then figure out the system later.
The reframe
The system is the brand. Build both together.
The system was built using Atomic Design methodology: colour tokens and type scales defined first, spacing rules locked next, then components assembled from those foundations. Nothing was designed in isolation. Every component inherits from the system rather than overriding it.
This allows the product to scale across iOS, Android, and web while maintaining a consistent point of view.
Explore the Conserva Design System
Product photography was produced through an AI image generation workflow built around a structured prompting framework. The framework specifies lighting conditions, texture qualities, compositional relationships, and colour constraints before any prompt is submitted. Without those constraints, AI-generated images produce variety where the product needs coherence. The prompting framework is, in effect, a visual brief applied to a new medium — design thinking about what the system needs, expressed as a specification for the tool.
Full-bleed imagery anchors categories and product views. Rather than competing with type or icons, photography does the cognitive work, reducing the need for labels and making visual richness the navigation mechanism, not decoration.
Colour appears sparingly and with purpose—calls to action, category anchors, key modules. It is never ambient.
Product descriptions open with sensory cues, Bright. Softened. Gently saline, before any technical detail. This sequence reflects how people actually decide whether they want something: feeling first, then fact.
A high-fidelity product where the point of view and the system that sustains it were designed as one thing, not two.
The prototype covers the full purchase flow—discovery, browsing, product detail, and cart—across platforms. The system defines tokens, components, and patterns at a level that can be carried into development without dilution. The AI prompting framework produced a visually coherent image library across product categories without manual policing of each output.
Test the sensory-first copywriting with actual users. The editorial logic (sensory cues before technical detail) is defensible, but it was never validated. Does "Bright. Softened. Gently saline." create confidence in a purchase, or does it create confusion about what the product actually is? I'd want to test comprehension and conversion intent against a more descriptive alternative before committing the pattern at scale.
Design for personalization within system constraints. The current system is deliberately consistent: the same tone, the same visual logic, everywhere. A more developed version might allow personalization (preferred categories surfaced differently, copywriting emphasis adjusted for return users) while maintaining the system's coherence. That's the harder design problem: how do you let a system adapt to individuals without losing the editorial point of view that makes it trustworthy?