GPT Image Generation Through Claude Code
The same thing as my Nano Banana app, but with OpenAI's GPT Image 2. I tell Claude Code what I want and it makes it. Public on GitHub with the skill.
This app is open on GitHub. Fork it, use it and build on it. I like giving it to the community.
About the project
A twin of my Nano Banana app, just pointed at OpenAI's image models: gpt-image-2 and its cheaper siblings.
I do it this way because every model is good at something different. GPT edits literally, it keeps the product exactly as it was in the reference photo and swaps the scene around it. Nano Banana, on the other hand, takes over the direction and fills in the art direction for me. When I need an e-commerce pack shot, I come here.
Pricing is per token, so the price list doesn't match reality. That's why after every call the app pulls the real cost from the API and tells me what it cost. In practice one image most often comes out at about €0.36, roughly three times Nano Banana.
Before paying, it can dry-run the job with the `validate` command, so a typo in a batch of twenty images gets caught for free. It also catches impossible combinations before you spend a cent: gpt-image-2, for example, can't do transparent backgrounds.
It's public on GitHub under the MIT license and comes bundled with a Claude Code skill that knows OpenAI's prices, limits and prompting rules.
What you get
- Literal photo editing: the product stays, the scene changes.
- The real cost from the API after every image, not an estimate.
- A free dry run of the batch before you pay.
- A cheap preview and an expensive final with one switch.
- Public on GitHub with the skill, fork it for yourself.
How it works (under the hood)
- gpt-image-2 and cheaper models (mini for previews).
- A bundled Claude Code skill, batches from JSON files.
- Up to 16 reference images, composited into a scene.
- Aspect ratios instead of pixels, output in PNG, WebP or JPEG.
Built with
Want to build like this too?
I teach building apps without programming in the AI First course.
More from Tool
Shoutbase: Testimonials Manager
An internal tool for managing testimonials: a queue of who to ask, a collection of what we've gathered, and an overview of where we use them.
Toucan: A Shared Second Brain for the Team
A shared cloud knowledge base that the team and its AI agents draw from, a "second brain" connected to my local context1.
Running a Project's Marketing Through Claude Code
A marketing team as a set of Claude Code skills: a director, department managers, and specialists that run a project alongside me.