How the AI works

A ONEH3RO report is written by an AI model, not by a person. Nobody reviews a report before you see it. The scores, the findings and the suggested fixes are all generated, and two runs on the same image can differ in wording and occasionally in score.

We think that is fine for what this is — a fast, consistent second opinion on artwork — and wrong for anything where you need a guarantee. Treat it as a well-briefed reviewer who is right most of the time, not as a measurement instrument.

Reports are produced by Anthropic's Claude family, called through the Anthropic API. The model receives a downscaled copy of your image — the same preview the report displays — together with a fixed rubric describing the nine categories and how to score them. It does not receive your email address, your account details or your purchase history.

Anthropic processes the image to generate the response. Under the Anthropic commercial API terms, inputs are not used to train its models. We do not use your images to train anything either, and we have no model of our own.

The rubric is about how an image reads: legibility at thumbnail size, contrast, focal hierarchy, product prominence, render quality, and compliance with Amazon's published image rules. Those are properties of the picture, which is why the same nine categories work whether you sell supplements or saucepans.

It is not connected to Amazon. It cannot see your listing, your sales, your conversion rate, your competitors' performance or your search ranking, and any statement it makes about how a change will affect sales is a designer's reasoning, not a prediction from data.

The markers on an annotated image are visual estimates, not detections. There is no object-detection model behind them. The box is the model's best guess at where something sits, and it can be off. Every marker carries its own written note, so use the note as the finding and the box only as a pointer — never crop, measure or compute anything from those coordinates.

Compliance flags are a prompt, not a ruling. Amazon's policies change, vary by category and marketplace, and are enforced by Amazon at their discretion. A clean report is not a guarantee your listing will be accepted, and a flag is not proof it will be suppressed. Check anything that matters against Amazon's current published policy for your category.

Scores are relative, not absolute. A 7 is a considered judgement against a rubric, not a measured quantity. Use the movement between runs on the same image, and the named fixes, rather than treating the number as precise.

You stay in charge of every decision. The report recommends; you choose what to change. Nothing here makes an automated decision that produces a legal effect for anyone, and no report is used to profile you.

If a report is plainly wrong or fails to complete, tell us at info@oneh3ro.com. A run that fails is refunded automatically — credits are only spent on a report that finishes.

The original file you uploaded is deleted automatically 28 days after analysis. The downscaled preview stays with the report for as long as you keep it. Full detail is on the privacy policy.

We may change which model version produces reports as better ones become available. When a change materially alters how reports read or score, we will say so here and update the date at the top of this page.