Hi Cursor Team,
I am a heavy user of Composer 2.5 and Agent mode. I love the tool, but I frequently run into a major workflow bottleneck when dragging image assets into the Composer input box.
Currently, when I drag an image from macOS Finder into the input box, Cursor automatically forces it to the top as a global Vision Input. However, when I drag other file formats (like .xlsx), they become inline text tokens/pills that follow the cursor text flow.
This behavior causes two critical issues in daily development:
1. Hard to describe/reference images in prompts
Because images are forced to the top, it is very difficult to refer to a specific image or structure the prompt naturally when mixing text, code, and multiple images. I cannot pin the image to a specific line of my text prompt.
2. Critical Issue: Asset Distortion and broken code references
When an image is treated as a Vision Input, Cursor seems to compress or process the image data for the model to “see”.
If my requirement is simply to tell the AI: “Help me replace the current background image with this new image asset,” the AI fails completely. It no longer sees the image as a raw file asset but as a processed visual context. As a result, it either attempts to output a broken processed image path/base64, or completely messes up the file reference in the generated code.
Expected Solution
I hope there can be a strict distinction between “Use this image for Vision analysis” and “Use this image as a raw File/Asset reference”.
- Ideal behavior: When holding a modifier key (e.g.,
Option/Alt) while dragging an image, or via a toggle, the image should be inserted as an inline file token (@image.png) just like regular files, without being forced into the vision preview at the top.
Here is a screenshot of the current mixed behavior for reference:
Thank you for the amazing tool, hope this UX improvement can be considered!