A skill that makes Claude plan AI edits to an existing photo the way a compositing lead would. You bring a picture that is nearly right and say what is wrong with it: the background, the shape, a cable in the way, the light. Claude reads how the photo was lit and shot, writes down what must never change, such as your label, a face or the dish as served, and splits the job into passes that each change one thing. For every pass it writes the mask and the prompt for your tool, whether that is Photoshop Generative Fill and Expand, ChatGPT or Gemini, then reads each result before the next pass begins.
Inside are four reference files. A tool sheet covers Photoshop, ChatGPT, the GPT Image API and Gemini as checked in October 2026: the models in each picker, render sizes, credit costs, mask formats and delivery sizes, with sources. A recipe file takes background swaps, new aspect ratios, removals, relighting and recolouring, with a table of how much of a frame gets generated for each change of shape. A third file sets mask margins, the check order and how to put the original pixels back. A worked example turns one kitchen photo for a fictional brand into a feed ad and a Story. Each pass uses the least invasive route that can do the job, and nothing goes into a picture that is not true of the product.
Claude cannot run the tools itself. You run each pass and send back the downloaded file, not a screenshot, because screenshots hide misspelt labels and seams. Editors that redraw the whole frame can still change letters on a label, so the skill keeps your product out of those redraws or puts the original pixels back afterwards. Tool menus change every few weeks, and your screen wins when it disagrees with the sheet. Turn on code execution so Claude can measure sizes, map what changed between versions, check label colours and export every size. Turn on web search so it can read a tool's current help page.
Claude skill · version 1.0
AI image editing
Checking your access.
In the download · 6 files
- SKILL.mdThe method Claude follows, in six stages: read the photo, plan the passes, mask, prompt, check each result, then finish and hand over.
- README.mdHow to install it in claude.ai or Claude Code, how to start, and the licence.
- references/worked-example.mdOne kitchen photo for a fictional brand turned into a feed ad and a Story through all six stages, with a failed pass fixed and one rolled back.
- references/tool-routes.mdWhich tool for which edit, and what Photoshop, ChatGPT, the GPT Image API and Gemini can do as checked in October 2026: models, render sizes, credits, mask formats, delivery sizes and sources.
- references/edit-recipes.mdBackground swaps, new aspect ratios with a table of how much gets generated, removals, relighting, recolouring, added objects and skies, pass by pass with prompts for each kind of tool.
- references/masks-and-checks.mdThe protect list, mask margins and edges, working sizes and tiles, putting the original back, measuring changes with code, the check order and grain matching.
Install it and begin
01Install it in Claude CodeTerminal mkdir -p ~/.claude/skillsunzip ~/Downloads/ai-image-editing.zip -d ~/.claude/skills/# Start it with /ai-image-editing. 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 editing skill. Here is my photo of [WHAT IS IN IT]. I need it for [WHERE IT WILL RUN] and I want to change [WHAT SHOULD CHANGE]. I have [YOUR TOOL AND PLAN]. Start with the edit sheet and the protect list.Attach the original file rather than a screenshot, and your brand colours if you have them. Turn on code execution so Claude can measure sizes and check your label colours between versions.
Worked example
One photo, edited without breaking it
- 01Read how the photo was shot before changing anything: light direction, softness, colour temperature, camera height and depth of field.Every new pixel has to agree with the old ones. A wall lit from the right behind a jar lit from the left looks pasted at any size.
- 02Write the protect list, with every word on the label spelt out letter by letter.Editors that redraw the whole frame quietly change letters, caps and fingers. A list written first is the only reliable way to catch it.
- 03Split the job into passes that each change one thing, in order: canvas, removals, background, light, then the other sizes.When two changes share a pass and the result fails, you cannot tell which instruction broke it. Removing first means later fills see a clean scene.
- 04Use the least invasive route for each pass. A crop or a colour layer comes before the Remove tool, and a fill inside a selection comes before a whole-image redraw.Each step up redraws more of the photo. A selection in Photoshop protects pixels physically; a chat editor only protects them in words.
- 05Before extending to a new shape, work out how much of the final frame would be generated. Over 40%, crop part of the way first.A 16:9 frame expanded to 9:16 is 68% invented. Cropping to 4:5 first brings that down to about 30%, and the seams land on plain wall instead of the product.
- 06Read every result in a fixed order: protected things, the change, seams at 100%, light and shadow, texture, then truth.A beautiful result with a misspelt label has failed. The order stops taste from hiding a fault, and three failures on the same fault mean it is time to change route.
- 07Finish by putting the original pixels back wherever nothing should have changed, matching the grain, and logging which areas are generated.The client gets full camera resolution where it matters, and you can answer honestly when someone asks what is real.
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.
- 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.
- 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.
- 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.
- 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.
- 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.
- Google AI StudioFree tier, paid above it
Where Gemini API keys, projects, billing and spend caps are made, with a playground for the models.
Bad at Free-tier prompts are used to improve Google's products, and the billing screens changed several times in 2026.
Why it matters
Ask a chat to edit a photo and it tends to change several things at once, redraw the whole frame and hand back a picture with a misspelt label and a softer face. Nobody can tell which instruction caused the damage. This skill makes Claude write down what must survive before anything else, change one thing per pass on the least invasive route, and check every result letter by letter before the next step.
- Type
- technique
- Period
- ongoing