A skill that makes Claude check AI-generated images the way a strict picture editor does before anything is published. It judges each image at the size people will see it, then runs six passes on full-size crops: hands and anatomy, words and logos, the product against photos of the real one, light and shadow, skin and surfaces, and brand fit. Every fault gets a grid square, plain words, a severity from S1 (stop) to S4 (polish) and how sure Claude is. Then it ranks the faults and writes the fix for each one: a retouch, a local edit prompt for your tool, a composite of the real asset, a new prompt, or a reason to shoot it for real.
Inside are five reference files. A fault atlas lists what goes wrong in each pass, where it hides and how to confirm it, with numbers such as the colour difference that makes a cap the wrong colour. A disclosure file covers product truth, invented customers and testimonials, AI labels on Meta, TikTok, YouTube, Google Shopping, Amazon and Etsy, New York's synthetic performer law, the EU AI Act and copyright, each with its source and the date it was checked. A fix playbook gives the route for every fault and each tool's edit warnings. A worked example and tested measuring code complete the set. Labels and logos are never redrawn by a model; the real file goes in.
Claude cannot make or edit the images, and it cannot tell whether an image is AI-generated, so it never claims to. It sees large uploads scaled down, so it asks for 100% crops of faces, hands, labels and text, and marks anything it could not see as unchecked. Photos of your real product matter most: without them product truth stays unchecked, which can stop an image shipping. Turn on code execution so Claude can cut crops, measure colour against the real pack, read the file's metadata and map every pixel an edit changed. The legal notes are a working summary with sources, not legal advice.
Claude skill · version 1.0
AI image critique
Checking your access.
In the download · 7 files
- SKILL.mdThe method Claude follows, in six stages: intake, first read, close inspection, truth and disclosure, ranked fixes, recheck and sign-off, with the S1 to S4 severity scale.
- README.mdHow to install it in claude.ai or Claude Code, how to start, and the licence.
- references/worked-example.mdA full critique of three launch images for a fictional skincare brand, covering all six stages: an invented testimonial removed, a real bottle composited in, and a sign-off.
- references/fault-atlas.mdThe faults to look for in each of the six passes (anatomy, text and marks, product truth, light and physics, skin and surfaces, brand fit), how to confirm them and their usual severity.
- references/disclosure-and-rights.mdProduct truth, invented people and testimonials, seals and marks, AI labels on Meta, TikTok, YouTube, Google Shopping, Amazon and Etsy, New York, California, EU and UK rules, provenance metadata and copyright, with sources and dates.
- references/fix-playbook.mdThe five fix routes (retouch, local edit, composite, regenerate, reshoot), how to write a local edit, fix recipes for every fault, each tool's edit features and warnings, and hand-retouch steps.
Install it and begin
01Install it in Claude CodeTerminal mkdir -p ~/.claude/skillsunzip ~/Downloads/ai-image-critique.zip -d ~/.claude/skills/# Start it with /ai-image-critique. If this is your first skill, type /reload-skills once.On claude.ai there is nothing to unzip: upload the zip as it is. Every step is on Install and use a Claude skill.
02Your first message to ClaudeText Use the ai image critique skill. These images are for [WHERE THEY WILL RUN] and will be seen at [SIZE, FOR EXAMPLE PHONE FEED OR 8 X 10 PRINT]. They were made in [YOUR TOOL]. We sell in [YOUR MARKETS]. Photos of the real product are attached. Tell me which images are usable and what to fix first.Attach full-size downloads, not screenshots, plus straight-on photos of the real product if it appears. Turn on code execution so Claude can cut crops, measure colour and read metadata.
Common mistakes
Where AI image checks go wrong
Instead ofJudging the image full screen on a laptop and calling it fine.
Do thisJudge it first at the size people will see it, about 400 px wide for a phone feed, then inspect faces, hands, labels and text in 100% crops.
Instead ofAsking the model to fix a misspelt label or a wonky logo.
Do thisRemove the generated words and composite the real label art or logo file. A model that drew the label wrong once will draw it wrong again in a new way.
Instead ofPassing the product because it looks close enough.
Do thisCompare it side by side with photos of the real product: shape within about 3%, the right closure, colour within about 2 dE00, and every word on the label.
Instead ofPutting a quote from an invented customer on the image and adding an AI label.
Do thisRemove the quote. A testimonial from someone who does not exist is a fake review under the FTC's rule, and no label fixes it.
Instead ofAsking whether people will be able to tell it is AI.
Do thisAsk what is wrong with it. Shadows that disagree, plastic skin and pseudo-text are faults in any picture, and naming them is what gets them fixed.
Instead ofStripping the metadata so the platform does not add an AI label.
Do thisKeep it. Google Shopping asks for an AI source tag in the file, Amazon asks for a keyword on photoreal AI people, and stripping provenance is never the fix.
Instead ofRunning five edits at once and trusting the result.
Do thisMake one change per pass, then recheck the whole frame, because generative edits can change pixels outside the selection and return a smaller file.
What to do it with
- ClaudeFree tier, paid above it
A language model that writes, edits, reads long documents and runs tools, including agents that work inside a codebase.
Bad at It is agreeable by default, so it will improve a bad idea instead of telling you it is bad. Ask it to argue against you and the quality of the answer changes.
- Claude CodeSubscription
An agent that reads a whole repository, edits files, runs the tests and explains what it changed.
Bad at It is fast enough to make a large mistake quickly. Work on a branch, read the diff, and never let it push. Watch the starting mode: on Pro, Max and Team plans it opens in auto mode, which Anthropic's own list says may read a .env file and send the key in it to the API that key belongs to. Switch it off.
- MidjourneySubscription
Generates images from a written description, with the strongest default aesthetic of anything in the category.
Bad at That default aesthetic is the problem as often as it is the draw. It has a house look, it pulls everything towards it, and getting a plain honest photograph of a real product out of it is harder than it should be. Text inside an image is unreliable.
- GPT ImagePay per use
Generates and edits images with unusually good instruction-following, and can hold text inside a picture better than most.
Bad at Faces drift between generations unless you feed a reference every time. Skin can come out plasticky at default settings, and a heavy fidelity setting costs many times more and takes many times longer.
- ChatGPTFree tier, paid above it
General assistant with image generation, data analysis and web access in one window.
Bad at Its default voice has become recognisable enough that readers now spot it, which is a brand problem if you ship it unedited.
- GeminiFree tier, paid above it
Google's assistant, with a very long context window and tight links to their other products.
Bad at Behaviour changes between the app, the API and the versions inside other Google products.
- Adobe FireflyFree tier, paid above it
Adobe's generative models, trained on licensed and public-domain material and wired into Photoshop and Illustrator.
Bad at Raw image quality trails the leaders. What you are buying is the indemnity and the fact that it sits inside files you already have.
- PhotoshopSubscription
Still the place a picture is finished. Generative fill, expand and remove now sit on the same layer stack as everything else.
Bad at Generative fill matches its surroundings rather than obeying a brief, so it is a patching tool and not a directing one.
- IdeogramFree tier, paid above it
An image model built around getting words right inside the picture.
Bad at Strong at type, weaker at photographic realism. A poster comes out well, a product on a plinth less so.
- FLUXOpen weights or open source
A family of image models, several with open weights, known for prompt adherence and clean type.
Bad at The open versions need a machine and a person who is comfortable with one. Licence terms differ between the variants, so check which one you are running before it goes anywhere near a client.
- C2PA Content CredentialsOpen weights or open source
A standard for attaching a signed record of how an image was made and edited.
Bad at Stripped by most platforms on upload, so its promise is far ahead of its reality.
- c2patoolOpen weights or open source
Reads, and can write, C2PA Content Credentials in images and video from the command line.
Bad at Reports Google's signing certificate as untrusted unless you configure a trust list; cannot see SynthID.
Why it matters
Ask a chat whether an AI image is fine and you tend to get praise and a vague worry about the hands. Meanwhile the label is misspelt, the cap is the wrong colour, the shadow falls the wrong way and the smiling customer with a quote does not exist. This skill makes Claude inspect in a fixed order, measure where it can, treat product truth and fake testimonials as a gate, and hand back a ranked list of fixes you can run.
- Type
- technique
- Period
- ongoing