Hi team,
Embedding visuals—whether game-screen mock-ups, prototype screenshots, or (in my case) side-by-side observed-vs-expected comparisons — makes feedback dramatically clearer to the LLM. You can drop each image path in square brackets, e.g. `[relative_path_to_your_screenshot.png]` in your conversation with Rovodev.
A single picture often saves a thousand tokens: Sonnet 4’s vision capability has been excellent at grasping expectations from these images, and I’d expect the same from the GPT-5.
If you try it and get good results, share your story so we can all benefit.
