Claude Prompts for Auditing AI Fashion Images Against Product References

Ask Claude to identify visible differences between a product reference and an AI-generated candidate, while marking anything it cannot assess. Use its response as a review aid, not a product-accuracy certificate or an automatic publishing decision.

A useful audit is specific enough to act on: “the visible collar points differ” is more helpful than “the image looks slightly wrong.” It should also be comfortable saying that a button count is unclear because the image is small or partly hidden.

The cover contains AI-generated shirt concepts. It is an illustration of a review workflow, not a validated before-and-after comparison or a photograph generated by Claude.

Prepare comparable evidence

Label the real source R and generated candidate G. Add matching detail crops when the original frames are too different to compare. Separate product references from moodboard pictures, so style inspiration is not mistaken for the garment specification.

Claude can analyse uploaded images, as described in Anthropic’s visual-capability guidance. That capability does not make a visual review equivalent to inspecting the physical product.

Eight copyable Claude fashion-image audit prompts

1. Check whether comparison is possible

Review reference R and candidate G. Before judging accuracy, identify whether their angle, crop, resolution and occlusion allow comparison. List features that can be assessed and features that need additional photographs. Do not treat an unseen feature as missing or assume that two different views prove a mismatch.

2. Build a visible garment-difference table

Compare collar, placket, pockets, sleeves and hem in R and G. Return feature, reference observation, candidate observation, visible difference and uncertainty. Use descriptive locations rather than invented pixel measurements. Mark matches cautiously and do not claim exact physical dimensions from photographs.

3. Inspect buttons without guessing hidden ones

Review the supplied placket close-ups. Count only clearly visible buttons and identify positions that are obscured or ambiguous. Check for duplicated buttons, uneven apparent spacing and a placket that bends unnaturally. Keep uncertainty separate from confirmed visible errors.

4. Review print or embroidery drift

Compare the artwork visible in both images. Look for changed motif scale, relocated decoration, mirrored motifs and invented elements. Anchor observations to collar, pocket, cuff or hem landmarks. Do not infer the real manufacturing technique or demand perfect seam matching unless the reference shows it.

5. Separate colour concern from colour measurement

Describe obvious visual colour or lighting differences between R and G. Distinguish a possible colour cast from a garment-design change. Do not assign colour-standard codes, numerical colour accuracy or fabric composition. List the controlled reference or additional photograph needed to investigate the concern.

6. Inspect anatomy and product visibility

Review G for visibly implausible hands, fingers, joints and garment intersections. Also flag hands, hair or props hiding important product details. Describe each issue’s location and why it needs human review. Do not identify the person or infer sensitive characteristics.

7. Turn findings into limited correction briefs

For each clearly supported issue, write one short correction instruction for an image editor. State what may change and which garment details should remain fixed. Keep uncertain observations as questions, not edit commands. Do not combine unrelated corrections into a single broad redesign request.

8. Prepare a human sign-off sheet

Create a review sheet from the audit: image ID, reference ID, confirmed visible issues, unresolved questions, required reference, proposed action and reviewer decision. Leave the reviewer decision blank. Do not declare the image approved, marketplace-compliant or factually accurate on the reviewer’s behalf.

Avoid a false sense of precision

A percentage score can look authoritative without being meaningful. Prefer documented observations and unresolved questions to an invented accuracy rating. Keep the original image, revised image and review notes together so the next reviewer can see what changed.

For a complete working sequence, first prepare the evidence using Claude garment-reference briefs, then consult TryBuy’s fashion QA guide before publication.

Frequently asked questions

Can Claude approve a catalogue image automatically?

This workflow deliberately leaves approval with a human reviewer. AI observations may be incomplete or incorrect.

What should I do when Claude flags a detail that looks correct?

Check the original at useful resolution and supply a clearer crop. Do not change a verified product detail merely because an AI review sounds confident.

Can the audit prove fit, fabric quality or authenticity?

No. Those claims require appropriate product evidence, not a comparison of generated pixels.

Platform reference checked 26 September 2026. Independent TryBuy guidance; no Anthropic endorsement or audit-accuracy guarantee is implied.