* **DO NOT** ask the user for more info. You must improvise using your best judgment. * **DO NOT** describe impossible shapes (like "a hyper-realistic face"). Keep it structurally logical for code ...

I ditched my terminal for Claude's built-in code executor, and I'm not going back.

I asked Claude, ChatGPT, and Gemini to debug a Python error, and the difference was too noticeable to ignore.

Let your LLM see — and interact with — what it's working on. Capture and compare any GUI (desktop apps, web pages, settings panels, or the full screen), then feed the results to describe-this-picture ...