photo-batch

---
name: photo-batch
description: Turns a folder of product photos into web-ready copies at the right sizes and formats, correctly named, with a report of anything that fails a shop's image rules. Originals are never touched.
---

# Every product photo renamed, resized and ready for the web

You point it at a folder of photos. You get back a new folder of web-ready copies at the sizes and formats a shop actually needs, named properly, plus a list saying which original became which file and which photos are too small to use.

Run it on your own machine, over your own files, in one pass:

```
codex exec -C "the folder" - < the-prompt.md
```

## What it does

1. **Open the source folder read only, and write everywhere else:** all output goes to a new `web/` folder created beside the source folder, never inside it. Before writing anything, record every original's file name, byte size and modified time in `photo-report.csv`, so check 2 can prove nothing was touched. If `web/` already exists and holds files, stop and say so rather than writing over them.
2. **Inventory every photo first:** one row per file in `photo-report.csv` with pixel width, pixel height, file size in MB, format, and the EXIF orientation value. Photos come off phones sideways, and orientation is the reason a batch of "correct" images ends up rotated on the shop.
3. **Fix rotation before resizing, not after:** apply the EXIF orientation and remove the tag. In Pillow that is `ImageOps.exif_transpose(im)`, documented as "If an image has an EXIF Orientation tag, other than 1, transpose the image accordingly, and remove the orientation data". Resizing a sideways photo first gives you a correctly rotated image with the wrong dimensions.
4. **Produce three sizes on the long edge, and never enlarge:** 2048 px for the main image, 1024 px for listing tiles, 512 px for thumbnails. Shopify's own help says "For square product images, a size of 2048 x 2048 px usually displays best". Pillow's `thumbnail()` is the safe call because it produces an image "no larger than the given size" and preserves the aspect ratio, so a small original stays small instead of being blown up into a blurry one. Every photo smaller than 2048 px on the long edge is listed in the report as needing a reshoot.
5. **Write WebP first and a JPEG beside it, and only add AVIF on request:** WebP is supported in "All versions of Chrome, Edge, Firefox, Opera, and Safari" according to MDN's image format guide, so it is safe as the main file. AVIF support starts at "Chrome 85, Edge 121, Opera 71, Firefox 93, and Safari 16.1", so the same page says you should include fallbacks using the `<picture>` element. If the shop cannot serve a `<picture>` element, do not produce AVIF at all: it is a file nobody will see.
6. **Use the encoder defaults unless the owner says otherwise, and never guess a quality number:** Pillow saves WebP at quality 80 and AVIF at quality 75 by default, and its WebP `method` defaults to 4 on a 0 to 6 speed against quality scale. Say in the report which numbers were used. If a saved file is bigger than the original, keep the original format for that photo and note it.
7. **Rename to a pattern the shop can sort:** `product-slug_<longedge>.<ext>`, lower case, hyphens instead of spaces, no accents, no brackets, no `#` and no `&`. Every output row in `photo-report.csv` names the original it came from, so nothing becomes unfindable. Never reuse an output name: if two originals produce the same slug, add `-2` and flag both in the report.
8. **Pass the colour profile through and strip location data:** Pillow only writes the ICC profile and EXIF block if you hand them to `save()`, and its documentation warns that "If a writer doesn't recognise an option, it is silently ignored", so a mistyped option leaves you with washed out colours and no error. Pass `icc_profile=im.info.get("icc_profile")` and do not carry GPS tags into a public file.
9. **Check each output against the shop's published rules and flag failures:** Google's Merchant Center requirements are at least 500 x 500 pixels, around 1500 x 1500 recommended, no image over 64 megapixels or 16 MB, and no watermarks, logos, pricing or calls to action laid over the picture. Shopify accepts up to 5000 x 5000 px or 25 megapixels and under 20 MB. Anything failing a rule gets a plain-English reason in the report, for example "too small for Google Shopping at 420 x 420".

## Then it checks

1. The file arithmetic balances: number of output files equals number of usable originals multiplied by the number of sizes multiplied by the number of formats, minus any skipped and named in the report. Print the sum.
2. Every original's byte size and modified time match what was recorded in step 1. If a single one has changed, an original was written to, and that is a failure whatever else worked.
3. No output is larger in pixels than the original it came from, and each output's width divided by its height matches the original's to within one per cent, which is what catches a squashed image.
4. Three outputs are opened and looked at: right way up, right subject, not stretched, colours not washed out.
5. Every output filename is unique, is lower case, contains no spaces and appears exactly once in `photo-report.csv` beside the original it came from.
6. Every photo listed as failing a shop rule has the specific rule and the actual measurement written next to it, not just the word "failed".

Any check fails: name it, redo that step once. Failed twice: say what is wrong and stop.

## Rules
- Public information only. This job needs no internet: the photos stay on the owner's machine, and nothing is uploaded to an online converter or compression service.
- Never invent a fact, a number or a quote. Do not state a compression saving you have not measured, and do not claim a size requirement that is not in the shop's own published rules.
- **Never overwrite an original, and never write inside the source folder.** If the only way to finish is to modify a source file, stop instead.
- Do not remove or add a watermark, a logo, or anyone else's branding. If the photos came from a supplier or a stock library, the licence decides what may be published, and this skill prepares the files so a professional can check that. It is not legal advice on image rights.
- Anything that cannot be produced is reported, never quietly skipped. A photo too small for the requested size is named in the report with its actual dimensions.

## Built from
- MDN Web Docs, "Image file type and format guide", https://developer.mozilla.org/en-US/docs/Web/Media/Guides/Formats/Image_types, last modified 28 August 2026: took the browser support statements for WebP and AVIF, and the instruction to include a `<picture>` fallback when using AVIF.
- web.dev, "Learn Images: WebP", https://web.dev/learn/images/webp, last updated 1 February 2023: took the fact that WebP uses a single quality value "expressed from 0-100, just like JPEG", which is why the skill sets one number and reports it rather than tuning per photo.
- Pillow documentation, https://pillow.readthedocs.io/en/stable/handbook/image-file-formats.html and https://pillow.readthedocs.io/en/stable/reference/ImageOps.html, read 7 September 2026: took the WebP quality default of 80, the AVIF quality default of 75, the WebP method default of 4, the behaviour of `thumbnail()` which never enlarges and preserves aspect ratio, `exif_transpose()` for orientation, and the warning that an unrecognised save option is silently ignored.
- Google Merchant Center Help, "Image requirements", https://support.google.com/merchants/answer/6324350, read 7 September 2026: took the 500 x 500 pixel minimum, the roughly 1500 x 1500 recommendation, the 64 megapixel and 16 MB caps, the ban on watermarks and promotional overlays, and the rule that the product fills between 75 and 90 per cent of the frame.
- Shopify Help Center, "Product media types", https://help.shopify.com/en/manual/products/product-media/product-media-types, read 7 September 2026: took the 2048 x 2048 px recommendation for square product images, the 5000 x 5000 px and 25 megapixel ceiling, and the 20 MB file size limit.