How I Edit My Photos with Claude

Written August 2026. Nothing here needs a design background, but it does need a Claude account and a folder of your own photos.

I'm rebuilding my brand and my website right now, and I needed photos that matched the direction I'm going. A real photoshoot isn't happening this month (two kids, one of them brand new).

So I opened the folder where all my old pictures live, which is about 30 of them going back 5 years. They were different outfits in different rooms on different cameras, and nothing about them looked like a set.

By the end of the afternoon I had 35 finished images that all look like they came from the same place, with the same treatment, color, grain and feel.

Here's exactly how I did it in 5 steps, and the last one matters most.

1. Go find examples of what you want

Open Google, search for the kind of photo you like, and save everything that stops you.

I searched things like "grainy film photo aesthetic" and "blurry flash photo" and "black and white editorial photo," and I saved about 8 images. There were grainy ones that look like security footage and blurry gold ones where the flash caught someone mid-turn. There was hard black and white with giant pink type across it, and washed-out cream covers with the subject off to one side.

Here's the part that felt wrong: I couldn't describe a single one of them. I just knew I liked them.

Save them anyway, because you're not going to copy them. You just need something to point at.

Where else to look: Pinterest works, and so does any Instagram account whose photos you keep saving. Mine came from one account I'd been collecting from for months, and it turned out every look I wanted was in one place.

2. Ask Claude to describe them back to you

This step is easy to skip, and it's the one that makes the rest work.

Claude can't copy a picture, so it needs the look written out as instructions before it can apply that look to your photos.

So upload your examples and ask it to describe them (describe is the word that matters there).

Here's the prompt I use:

Here are the images I've been saving. For each one, describe what was done to the photograph, written as instructions someone could actually follow.

Cover all of these:
- Color and contrast. How dark are the blacks, how blown out are the whites, what color cast is on it.
- Framing. Where is the subject in the frame, what got cropped off, what is out of focus.
- Texture. Grain, blur, scan lines, anything sitting on top of the image.
- Type, if there's text on it. Size, color, placement, whether it's cut off.

Do not tell me it looks moody or cinematic or editorial. Those words don't help me. Tell me what was physically done to the picture.

The last line is doing the most work, because without it you get adjectives, and you can't build anything out of an adjective.

What came back for the security-footage look:

Black and white, contrast pushed until skin blows out and shadows go solid. Four separate camera feeds in a grid with thin black gaps. Every crop is deliberately wrong. Chins cut off, heads half out of frame, nothing centered. Scan lines and grain over everything. Small monospace text in the corners: REC, a timecode, a battery icon.

There's something in that description I'd never have found on my own.

The crops are what make it work. If I'd only copied the grainy filter part and put it over a nice centered portrait, I'd have gotten a photo with a filter on it. Cutting the chin off is what makes it look like a camera that didn't know it was recording you.

That's what a description gets you that copying never does, because it finds the rule underneath.

Push back if it's vague. If it says "warm tones," ask warm how (toward orange or toward yellow?). If it says "high contrast," ask whether the blacks still have detail in them. You want it specific enough that you'd know if it was wrong.

3. Hand over your own photos

Now give it your pictures and ask it to apply one of the descriptions.

This is where I got told no, and I'm glad I was.

I asked it to use an AI image editor on my photos, which I'd already tried a few months back on my headshots. What came back was a woman with my haircut, flat dead eyes and a melted hand wrapped around a coffee mug. I put it on my homepage, looked at it and said, this is so bad.

Here's what I didn't know: Claude can look at a picture and tell you everything about it, but it can't make one, because there's no image generator inside it yet. So when I asked for one, it didn't pretend.

It wrote code instead.

The code is Python that opens my actual photo file, works out where my face and the light really are, crops it and then adjusts the color. My real photograph goes in and my real photograph comes out treated, and since nothing is generated, I still look like me.

Original photo of Christi in a denim jacket, sharp and in full color — This is the same picture, taken years ago. The crop cuts my head off at the chin so it's about the hands instead of my face, and then the color goes warm and a directional blur runs across the frame like the camera moved. **It's still my hands and my jacket.** Nothing here was drawn by an AI.

The same photo cropped at the chin, motion blurred and graded warm gold — This is the same picture, taken years ago. The crop cuts my head off at the chin so it's about the hands instead of my face, and then the color goes warm and a directional blur runs across the frame like the camera moved. **It's still my hands and my jacket.** Nothing here was drawn by an AI.

The rule I'd give anyone: never point an AI image generator at a real person you care about, or at anything that's supposed to be a true record. A real photo that's a bit off for your layout beats a generated one that's perfect, every single time.

One thing worth asking for. Before it crops anything, have it measure the photo. Mine builds a version of the picture with a grid over it, reads off where the face sits and crops from those numbers. When I let it guess instead, it put a crop in the wrong place four times in a row, and measuring took one try.

4. Tell it what's wrong

This is most of the work, and it's the part I'd have skipped a year ago because it feels like the project going badly, when really it's the project working.

Three things went wrong for me.

My kids' faces. I'd asked for them to be blurred out, and what came back was a small child with a smeared-over face that looked like a horror movie. I said so, and now it crops them out of the frame instead, so you get their hands, their feet, a toy truck or the back of a head. Those turned out to be better pictures anyway.

The color was too yellow. I said that too, and it went and looked at the account I'd been saving from, worked out that the palette is rust and cognac rather than olive green, and fixed the setting properly instead of nudging it.

It only made one of my favorite ones. When I asked why, it turned out it had applied a rule from a different style where it didn't belong, so it built three more.

Say what you don't like in normal words, because "That's creepy" is a complete piece of feedback. Naming what's wrong is enough, and working out the fix is its job.

5. Save the setup as a skill

Everything above gets you good photos once, and this step is what makes it yours.

A skill is a folder on your computer with a set of written instructions inside, plus any code that goes with it, and Claude reads it before it starts working. You make one by asking:

Save all of this as a skill so I don't have to set it up again. Put the descriptions of each style in it, the code you wrote, and every correction I made today with the reason behind it.

That's it, and it writes the folder for you.

Here's why it's the step that matters. A chat is gone tomorrow, so everything I explained (every correction, every "no, not like that") disappears when the conversation ends. Next month I'd start over and re-explain it to a version of Claude that had never met me.

A skill keeps it, and mine now holds these rules on its own:

Never generate my face. Only treat real photographs.
Never smear out a kid's face. Crop them out of the frame instead.
The color is rust, not olive.
Measure a photo before cropping it and never guess.
Ship the picture clean with no text on it. I add my own.

I didn't have to write any of that down. I just had to notice it while it was happening and say it once.

What I actually ended up with

I put 31 photos in and got 35 finished images back, in 5 looks that all belong together:

The security-camera grid. Four shots in one frame, black and white, timecodes in the corners. It's my favorite of the set and the one that looks least like a filter.
The blurry flash one. It's warm gold and motion smeared, cropped at the neck so it's about the hands.
Hard black and white. The blacks have no detail left in them at all, and the whites are blown to paper.
A washed-out background with one side left empty, so I can drop type on it later.
Warm film color. It's sharp, rust and cream, and it's the one I used for most of them.

That's enough to rebuild the site, out of pictures that were already on my hard drive.

And I still can't edit a photo.

The setup, if you want to run this yourself

All of this happened in Claude Code, which is Claude running in your terminal instead of a chat window, and that's the part that matters. A chat window can't reach the photos on your hard drive, but the terminal can, so Claude can open your actual files, write the code, run it, look at what came out and fix it.

If the word terminal makes you want to close this tab, stay with me. You type a sentence in plain English and it does the work, so you aren't writing code. You're just working in the room where your files already live.

Where nano banana comes in

Nano banana is the nickname everyone uses for Google's image model. It's the one that actually makes and edits pictures, and it's very good at props and texture: paper stock, tape, torn edges, backgrounds, objects on a desk.

You can try it in the Gemini app at gemini.google.com. To use it inside Claude Code you need a free API key from Google AI Studio, and Google's own notes on it are here.

So here's the split in my setup. Claude reads my photos, writes the code and runs it, and nano banana makes the paper and the tape. Nano banana never touches my face, or my kids' faces. It's props and texture only, and that rule came out of the melted hand.

Plugging one into the other

MCP stands for Model Context Protocol, and it's how Claude Code talks to outside tools. Think of it as a plug: somebody writes the adapter once, you plug it in, and Claude can use that tool the same way it uses everything else it has.

The one I use is nanobanana-mcp, so download it and install its dependencies:

git clone https://github.com/charlesdove977/nanobanana-mcp
cd nanobanana-mcp
npm install

Then tell Claude Code it exists with one line, which you can run from anywhere:

claude mcp add nano-banana \
  -e GEMINI_API_KEY=your-key-here \
  -- node /full/path/to/nanobanana-mcp/dist/index.js

Swap in your own key and your own path, then open Claude Code and type /mcp to check it connected. Once it's in you don't think about it again, and when you ask for a paper texture you get one.