GPT-Image-2 Ecommerce Asset Generator

Problem

E-commerce image generation often suffers from unstable input rather than missing model power: product selling points, category, scene, style, size, and reference images are scattered across the request. Stacking those fragments into a prompt can produce generic, over-polished, or visually inconsistent product images. This skill turns a generation request into a structured workflow, matching templates and producing a concise prompt for GPT-Image-2.

How it works

• Intent recognition: extracts scene type, product category, style preference, and optional reference image path, such as hero image, flat lay, ugc, packaging, or device mockup.
• Template matching: reads one matching template under references/templates/ based on trigger words, falling back to 01-hero-image.json when no template matches, instead of loading every template.
• Prompt assembly: uses prompt_template, substitutes variables, applies variants and category_tips, removes empty fields, and outputs a short, focused JSON object rather than a long keyword list.
• Generation and cleanup: invokes image generation through Codex CLI or a local HTTP service, copies the image to the working directory, renames it, and deletes the original session folder under ~/.codex/generated_images/ to avoid duplicate storage.
• Scope: works best for templated e-commerce assets such as hero images, lifestyle scenes, posters, model shots, detail pages, and social media visuals. For abstract concepts or productless creative scenes, more context is still needed. For UGC, livestream, and social content, it adds concrete phone models, grain, warm casts, and natural framing to reduce over-retouched looks.