Make Images with Claude Code

Make an image for me with Codex.

You cannot make images yourself. The ChatGPT extension for VS Code comes with Codex, a command line tool that can. Use it like this.

1. Find Codex. Try codex on PATH first (codex.cmd too on Windows; on a Mac also check /opt/homebrew/bin and /usr/local/bin). If it is not there, look under ~/.vscode/extensions, ~/.cursor/extensions, ~/.vscode-insiders/extensions, and ~/.vscode-server/extensions (WSL and remote) for a folder that matches openai.chatgpt-*. Inside it the binary is bin/windows-x86_64/codex.exe on Windows, bin/macos-aarch64/codex or bin/macos-x86_64/codex on Mac, and bin/linux-x86_64/codex on Linux. Ignore the other bin folders, they are for other systems. Two versions side by side: take the highest version number, compared number by number, not alphabetically. If you find nothing, stop and tell me to install the ChatGPT extension in VS Code (marketplace id openai.chatgpt), sign in with my ChatGPT account, and reload the window. Do not install anything yourself. No npm, no downloads. Wait for me.

2. If I have not said what the image is, ask me: the subject, the setting and light, the framing, the style, and what to leave out. Write that as one clear brief. The brief is the scene only. File names and folders stay out of it: the folder is where you run Codex, and the file name goes in the command.

3. Run Codex from the folder where the image should be saved. Create the folder if it is missing, and do the cd in the same command as the run. CODEX is the path you found in step 1, in quotes. One image per run:
   cd FOLDER && "CODEX" exec --sandbox workspace-write --skip-git-repo-check "Generate one image saved as NAME.png in the current working directory: BRIEF" < /dev/null
   Use a lowercase file name with hyphens, no folder in it. Never overwrite an image that exists; add -v2 instead. Before you run it, note the time and whether a file with that name already exists. A run takes 2 to 3 minutes, longer than your default command timeout, so give the command a 10 minute timeout (600000 ms) or start it in the background and check the folder when it ends. --sandbox workspace-write lets Codex run without asking and write files only in the folder it starts in, so start it in the image folder, never at the root of my drive. --skip-git-repo-check lets it run in a folder that is not a git repo. Do not use --full-auto: Codex removed that flag in version 0.154 and the run fails before it starts. PowerShell: put & before the quoted path, use ; between the commands, and drop the < /dev/null part.
   Several images: start one background run per image at the same time, up to six, each with its own file name, then collect the results.

4. Check the file yourself. It must exist, be bigger than 0 bytes, and have a modified time later than the time you noted. Do not trust Codex saying it saved something. Then look at it and say what you see: count hands, fingers, and objects, and read any lettering. Flag anything wrong before I ask. Then show me the image.

5. Report one line: OK with the path and size, or FAIL with the last 10 lines of the Codex output.

To edit an image I already have, keep the prompt BEFORE the -i flag:
   "CODEX" exec --sandbox workspace-write --skip-git-repo-check "Keep everything exactly as it is. Change only WHAT. Save the edited image as NAME-v2.png in the current working directory. Do not overwrite the input file." -i NAME.png

If something goes wrong, do this before you ask me:

- Your permission system blocks the Codex run (the message says a classifier blocked it): stop. Do not work around it. Tell me to press Shift+Tab until the line under the prompt box no longer says auto mode, then say go. Run the same command again and wait for me to approve it.
- Codex says "not logged in" or "usage limit": stop. Tell me to sign in to the ChatGPT extension again, or to wait for the limit to reset. Keep my brief so we can rerun it.
- The command stopped at 2 minutes with no file: that was your own timeout, not Codex. Look for the file first. If it is not there, run the same command again with the 10 minute timeout.
- The run ends but no new file appears: run the same command once more. If there is still no file, list the newest PNG files in the folder. Codex will say it saved something when nothing hit the disk. If a PNG written during this run shows up under another name, rename it to the one I asked for and carry on. If a folder was in the file name, drop it. Codex only writes into the folder it runs in.
- A background run is still going after 10 minutes: stop it. ChatGPT is slow or queued. Tell me, and try once more later.
- An edit fails with "no prompt provided": -i read the prompt as a file name. Put the prompt before -i and run again.
- Codex says its sandbox could not start (a Linux box, a container, or WSL): run the same command again with --sandbox danger-full-access in place of --sandbox workspace-write. Codex may then write anywhere, so keep the run inside the image folder.
- Codex says the flag --full-auto is unknown: the command still has the old flag. Use --sandbox workspace-write --skip-git-repo-check instead.
- I asked for a transparent background: you cannot tell real transparency by looking, and a checkerboard in a viewer usually means it worked. Check the PNG has an alpha channel and a corner pixel that is fully transparent (with Pillow: python -c "from PIL import Image; im=Image.open('NAME.png'); print(im.mode, im.getpixel((0,0)))"). Only if the background is painted in, change the brief to "isolated on a fully TRANSPARENT background, PNG with alpha, no backdrop, no floor, no shadow plate" and run once. No Pillow: say you could not check, do not guess.
- The image has extra fingers, doubled objects, or wrong text: tell me what you see. Rerun with that object removed from the brief ("no hands", "no cutlery") instead of counting it. Keep any text in the image to a few short words.
- After two edits on the same image, stop editing. Put everything we learned into one fresh brief and make it new.

Make me the hero image for my pizza page.

Subject: a wood-fired margherita pizza on a dark slate board.
Setting and light: a rustic oak table, warm side light from a window on the left.
Framing: close three-quarter view, the board fills the lower two thirds.
Style: editorial food photography, shallow depth of field.
Leave out: text, hands, cutlery.

Save it in assets/img as pizza-hero.png.