How to Write Prompts for AI Product Photography: The Complete Guide

By the ORA Lab team · Updated 29 August 2026 · 10 min read

The worst product prompts read like reviews of an image that doesn't exist yet: 'stunning professional photo of luxury ring, ultra realistic, 8k, masterpiece.' Every word is a compliment; none is an instruction. The model, given nothing to work with, falls back on the averages of every ring photo it has seen, and you get the generic glossy output that reads as AI at a glance. The best product prompts read like a brief to a photographer who can't ask follow-up questions: here's the scene, here's the surface, here's where the light comes from, here's the lens. Specific, visual, and silent about quality, quality is the tool's job.

This guide is the prompt-writing method we teach brand teams, written to transfer across tools: the anatomy of a complete prompt, the vocabulary that does the steering, the revision discipline that protects your keeper rate, and the boundary, which matters more in product work than anywhere else, between what prompts should control and what they must never be trusted with.

First principle: the prompt directs the scene, the photo anchors the product

In product-anchored generation, your clean source photo carries the product's identity, geometry, materials, stone map, and the prompt's job is everything around it. This division of labour is the whole game. The moment your prompt starts asserting product facts ('emerald-cut solitaire, six prongs, rose gold band'), one of two things is true: either the tool is anchoring the product properly and the assertions are redundant, or it isn't and they're a prayer. Tools that need the prompt to describe the product are regenerating it from averages, which is disqualifying for commerce regardless of prompt skill. Write prompts assuming the product is handled; if it isn't, no prompt fixes that.

The six-layer anatomy

Assembled: 'Centered on a low pedestal at a three-quarter angle, in a minimalist gallery interior at early morning. Honed travertine surface in warm ivory. Soft window light from the upper left with gentle falloff, no harsh speculars. Macro distance, 100mm feel, shallow depth of field, camera at product height. Muted warm grade, low contrast, editorial catalog finish.' Fifty words, zero adjectives about quality, and a photographer could shoot it tomorrow, which is exactly the test a prompt should pass. When output disappoints, diagnose by layer: flat image, look at lighting; wrong mood, look at grade and setting; awkward composition, look at camera and product context. Missing layers are found this way faster than by re-rolling.

The vocabulary that steers

• Light by direction and character, not adjective: 'backlit through sheer linen', 'hard noon sun with crisp shadows', 'single softbox from camera left', each produces a different, predictable image; 'dramatic lighting' produces a lottery.
• Borrow lens language even in generation: '85mm portrait feel' compresses and flatters; 'wide 35mm environmental shot' shows room context; 'macro' signals detail scale. Models have internalised what photographs taken these ways look like.
• Name surfaces precisely: 'brushed concrete', 'raw silk', 'wet slate', 'aged brass tray', surface words are the highest-value words in product prompting because the surface fills most of the frame that isn't product.
• Grade with references, not superlatives: 'warm editorial grade', 'high-key catalog white', 'moody low-key with deep shadows', the vocabulary of retouching briefs transfers directly.
• Cut the incantations: '8k', 'ultra-realistic', 'masterpiece', 'trending' add nothing in modern systems, they're folk magic from 2022 that survives by superstition, and they crowd out words that steer.

Revision discipline: direct, don't gamble

The difference between teams with 60% keeper rates and teams with 20% is rarely the first prompt, it's what happens after a near-miss. The gambler's response is to rewrite everything and roll again, which discards the four layers that worked to fix the one that didn't. The director's response is surgical: keep the prompt, change one layer, regenerate. 'Same scene, light from the right instead.' 'Same everything, pull the camera back.' One variable per iteration, so every result teaches you something about your tool's dialect. Keep a prompt log, the exact wording of every keeper, pasted into a shared doc, because a validated prompt reused across twenty SKUs is how batch consistency happens without a style guide meeting, and how the same-scene-different-product test becomes your daily workflow instead of an evaluation exercise.

Beyond prompts: where this is all heading

A candid note on the craft you've just learned: the industry is actively engineering it away. The six-layer anatomy exists because raw generation models need every decision spelled out per image, but brand imagery doesn't actually vary six ways per shot. Your surfaces, grades, and lighting characters should be brand decisions made once, leaving only the per-shot choices to specify. That's the direction serious tools are moving, locked visual systems plus lightweight per-shot direction ('hints' rather than paragraphs), which is the model ORA works on and the subject of the next piece in this series. Learn prompting properly, it's the literacy layer, and it makes you a sharper art director in any tool, but treat 200-word prompt recipes as scaffolding, not the destination.

If you'd rather skip to the destination, brand look locked once, per-shot direction in a sentence, book a demo and bring your three hardest SKUs plus the moodiest reference image on your board. Watching your own product land in that scene without a paragraph of prompt is the fastest way to understand the difference.