- **Learn from each step:** Using mechanisms like internal prompt self-critique to refine its own instructions for the image generation model. - **Adapt the description (optionally):** Allowing text co-iteration to clarify the textual input for better visual output. - **Continuously self-evaluate and correct:** By repeatedly trying to meet the strict criteria for patent drawings, including eliminating extraneous text, ensuring correct styling (like black lines on a pure white background), and improving the visual clarity of specified labels and figures.