Case Study · AI Tooling

Prompt Optimizer

A prompt-engineering tool I built for design teams: it reads a brand's inspiration images and writes the prompt language that keeps every AI generation on-brand.

← Back to Skills ← Back to Work Samples

The Problem

AI image tools don't drift. Prompts do.

When a design team starts generating imagery with AI, every designer describes the brand a little differently. One writes "soft natural light," another writes "bright and airy," a third writes nothing about light at all. Fifty generations later, the brand's visual identity has quietly dissolved into fifty interpretations of it.

The fix isn't a better model — it's better language. A prompt is a brand guideline written for a machine, and it deserves the same rigor as one written for people. Prompt Optimizer automates the hardest part: translating what a brand's photography looks like into words a model can reliably obey.

How It Works

Brief. Anchor. Synthesize.

01 / Brief

The designer types the ask in plain words. "birkenstock mommy and me image" is a complete input — the engine does the rest.

02 / Anchor

Two reference images ground the output: Product DNA — the exact product that must appear, down to the buckle — and Aesthetic Inspiration, the shot that sets mood, pose, and palette.

03 / Synthesize

The engine writes the full production spec: cast, pose, environment, lighting, capture quality, hard constraints, and a render-ready directive.

The Prompt Synthesis Engine workbench: a three-word artistic directive and two reference images on the left, and the full synthesized spec — visibility, fidelity, pose, cast, environment, lighting, constraints, and directive — on the right — The full workbench. Left rail: a three-word directive, the Product DNA image, and the aesthetic inspiration. Main panel: the generated spec — visibility and fidelity rules, pose, cast, environment, lighting engine, four hard constraints, and the final directive. Scroll to see the full width on smaller screens.

A Closer Look

How the brief steers the prompt

Close-up of the input rail: the three-word artistic directive 'birkenstock mommy and me image', the Product DNA reference thumbnail, and the Aesthetic Inspiration reference thumbnail — The Brief Three words, plus two reference images. That's the entire input.

Close-up of the synthesized visibility and fidelity rules: branding must stay sharp and legible, and the sandal silhouette must remain anatomically correct — Visibility & Fidelity What must stay sharp, legible, and anatomically correct — pulled straight from the Product DNA image.

Close-up of the synthesized pose and cast: the mother seated in a rattan chair, daughter standing beside her, and the specific cast description — Pose & Cast The exact pose and cast, mirrored from the aesthetic inspiration image.

Close-up of the synthesized primary setting: an outdoor patio with warm-toned stucco walls, a large stone archway, and a weathered wooden door backdrop — Setting The scene, described with production-brief specificity.

Close-up of the synthesized lighting engine: golden hour natural light, creating soft directional shadows and a warm, high-key glow — Lighting Mood, translated into a lighting setup a photographer could actually rig.

Close-up of the synthesized capture quality: photorealistic, commercial-grade editorial photography — Capture Quality The finish the render is held to.

The Copy Problem At The Core

From taste to language

This is where the writing discipline earns its keep. Here's a real run — what went in, and what the engine wrote:

The entire input — what the designer typed birkenstock mommy and me image

Three words, plus two reference images: the Mayari sandal (Product DNA) and one inspiration shot. That's the whole brief.

The synthesized directive — what the model receives

Commercial lifestyle photography of a mother and her young daughter in a 'Mommy and Me' moment. Both are wearing identical Birkenstock Mayari sandals in Mocha nubuck leather, featuring the distinct criss-cross toe loop, dual adjustable straps with bronze buckles, and anatomical cork footbeds. The mother is seated in a rattan chair against a warm stucco wall and an arched wooden door; the daughter stands next to her. Both wear white textured cotton dresses. Lighting is warm golden hour sunlight with soft shadows, creating a Mediterranean summer vibe. 8k resolution, shot on Hasselblad H6D, 50mm lens, f/2.8, sharp focus on the sandals, natural textures of leather and cork, highly detailed skin and fabric textures, photorealistic.

Every clause is traceable to one of the two references: the sandal's construction comes from the Product DNA image, the pose, wardrobe, and setting from the inspiration shot. Nothing is left for the model to guess.

The guardrails — synthesized alongside the directive

Do NOT use the footwear from the inspiration images (heels/flats)
Only the Mayari sandal from the Product DNA image is permitted
Maintain the nubuck texture and cork grain without AI-generated smoothing
No text or graphic overlays

The Proof

References in. Finished shot out.

The directive above produced the image below. The exact Mayari — criss-cross toe loop, bronze buckles, mocha nubuck — on both models, in the inspiration shot's world: white cotton dresses, rattan chair, warm stucco, golden hour. The inspiration image's heels never made it through, because the constraints said they couldn't.

Aesthetic inspiration reference: a mother seated and daughter standing in white dresses in front of an arched wooden door — Input · Aesthetic Inspiration Sets pose, wardrobe, palette, and mood.

Product DNA reference: the Birkenstock Mayari sandal in mocha nubuck — Input · Product DNA The exact product that must appear — down to the buckle.

The generated image: mother seated in a rattan chair and daughter standing beside her, both in white dresses wearing the Birkenstock Mayari sandals in a golden-hour Mediterranean courtyard — Output · Generation The synthesized directive, rendered: the inspiration's world, wearing the Product DNA's sandal.

Built with React · Vite · Gemini multimodal analysis · structured JSON output

All screenshots and imagery show a real synthesis run from brief to render.

← Back to Work Samples