Imagine teaching a computer to see and understand the world, not by writing complex code, but by showing it a few examples. That's the promise of visual programming, a field where AI models learn to generate programs that solve visual tasks like answering questions about images. Traditionally, this has involved prompting large language models (LLMs) with detailed instructions and examples, a process that's both expensive and slow. But what if we could skip the prompting altogether? New research explores how to create visual programs without prompting LLMs during inference, potentially unlocking faster, cheaper, and more scalable visual reasoning. The key idea is to break down programs into reusable building blocks called 'templates' and their corresponding 'arguments.' For example, "Count the red chairs" and "Count the green bananas" share the same template (counting) but have different arguments (object color and type). This allows researchers to create synthetic data by simply swapping arguments, training smaller, more efficient LLMs to generate visual programs on the fly. This prompt-free approach shows promising results, achieving comparable performance to traditional methods on visual question answering tasks while significantly reducing cost and inference time. This could open doors to wider adoption of visual programming, enabling applications like image search, robot navigation, and accessibility tools for the visually impaired. While challenges remain, such as handling complex visual scenarios and improving the accuracy of visual models, this research points towards a future where AI can learn to "see and code" more efficiently than ever before.