Spellcast Universe

Inspiration and What I Learned

Since the launch of GPT Image 2, I have spent a great deal of time studying the model and experimenting with how it generates images. While working with anime-style imagery, I found that Image 2 often produced excessive visual noise. This became the starting point of my research into prompt engineering.

I began by structuring and modularizing prompts, and developed a technique I call the Noise Reduction Ladder. I then continued studying many different kinds of image-generation prompts in greater depth.

One of the most important things I learned is that prompts are not a universal language. A prompt that works for one model may not work the same way for another. For that reason, my research has been focused specifically on prompt engineering for OpenAI GPT Image 2.

I realized that the power of Lisp macros could turn one stable prompt into 100, 1,000, or even 10,000 controlled variations. Lisp can manage the boundaries and perform symbolic compilation, while Codex brings creativity into the structure. To me, this is the best combination: Lisp provides control and scalability, and Codex provides imagination.

That idea led me to create Prompt Universe.

How I Built It

Prompt Universe treats prompts as structured and evolvable programs rather than disposable blocks of text. A stable prompt becomes a reusable seed. Lisp macros can then expand that seed across different creative dimensions while preserving its important rules, structure, and visual identity.

After building the ability to evolve a single spell into a large family of prompts, only one major problem remained: generating all of the images quickly.

I spent a great deal of time studying Codex and its imagegen skill. I then built local scripts that allow Codex to control multiple Terminal windows and invoke additional Codex workers in parallel. The jobs are divided into controlled waves so that image generation can be accelerated without colliding with rate limits.

In other words, the system makes the fullest possible use of Codex's image-generation capacity within each available time window. It is highly efficient, but it is also practical for ordinary users who want to generate many creative variations without manually repeating the same workflow.

I called this tool image-gen-runner, and it is also being prepared for release.

Challenges

The first challenge was understanding why anime-style generations from GPT Image 2 could become noisy or visually fragmented. Solving that problem required much more than adding negative keywords. I had to study prompt hierarchy, material descriptions, semantic conflicts, attention priority, and the relationship between structure and visual stability. That research eventually became the Noise Reduction Ladder and the broader philosophy behind Prompt Universe.

The second challenge was scale. Generating hundreds or thousands of useful prompt variations is only valuable if those prompts can also be turned into images efficiently. Building image-gen-runner required me to study how Codex handles image generation, how to coordinate multiple workers, how to divide jobs safely, and how to avoid wasting capacity or overwhelming rate limits.

The final challenge was bringing these two very different systems together. Prompt Universe focuses on creativity, symbolic structure, compilation, and controlled evolution. image-gen-runner focuses on orchestration, parallelism, and generation efficiency. They were originally separate tools with separate workflows.

Bringing It Together

Participating in OpenAI Build Week gave me an opportunity to rethink both projects.

Instead of keeping Prompt Universe and image-gen-runner as separate tools, I decided to integrate them into a single CLI. The result is Spellcast Universe: a tool that combines Codex creativity, Lisp-powered symbolic compilation, scalable prompt evolution, and parallel image generation in one workflow.

I believe this makes the original ideas more complete and much more useful. A user can begin with a creative idea, turn it into a structured spell, evolve it into many controlled prompts, and then automatically cast those prompts into images through Codex—all with both creativity and efficiency.

Built With

• codex
• image-2
• lisp
• python