Descriptions written from the photograph

What it writes

Two short paragraphs in plain Indian English, describing colour, material appearance, silhouette, closure, sole and visible branding, plus the sizes you stock. No headings, no bullet points, no marketing filler.

Nike Free RN Flyknit Crimson

UK 6–11 · one photograph

Nike Free RN Flyknit in Crimson. The upper is a knit textile in red with a white Swoosh, dark burgundy lacing, and a burgundy pull tab at the heel. The sole is a bold pink-red with Nike branding visible on the side. A streamlined running shoe, available in UK sizes 6 to 11.

What it will not write

The failure mode for AI product copy is confident invention. Given one photo of a shoe, these models will happily produce “breathable mesh upper, waterproof membrane, responsive foam midsole” — none of which they can see. A customer buys on that and returns the shoe, and a return costs a small shop the shipping both ways plus the sale.

So there are two defences, and the second is the one that holds:

• The prompt forbids it at length, with concrete banned examples rather than abstract rules.
• The output is then checked in code against about two dozen phrases a photograph cannot support. If any appear, the model gets one corrective pass naming them. Anything that survives comes back flagged.

A prompt is guidance. A list is enforcement. Both were built because the first live test claimed a “gusseted tongue” on a shoe it had seen one photo of.

No photograph? A shorter answer

If you send no images, you get the brand, the model name and the sizes — and nothing about appearance. Describing a look it has not seen would mean recalling what that model name usually looks like, which may be a completely different colourway from the one on your shelf.

Doing a whole catalogue

Up to 200 products in one job, at the same price per description. It runs in the background and you are charged only for the ones that come back — forty submitted and thirty-eight written costs thirty-eight.