Make the first competent request to the “sighted” AI using this photo. Return JSON {"prompt":"...","goal":"...","context":"..."}: prompt — what exactly are you ASKING the AI (the question itself, for example, “what does this icon mean?”), goal — what RESULT you want to get as an output (this is different than a question: for example, a clear explanation and conclusion, is it possible drive safely), context - what is important to know AI, what is not visible in the photo (make/model of the car, when it caught fire, etc.).
“AI sees” is not magic. The model splits the image into thousands of small sections and compares them with what it “remembered” from millions of images: shapes, text, objects, colors. The result is that it describes the picture with words inside itself, and then works as with ordinary text. There are two conclusions from this. First: the clearer the picture (sharpness, light, the entire object in the frame), the more accurate the answer - like a person in poor lighting. Second: AI can confidently make a mistake in a small detail, so the picture should not just be “thrown”, but accompanied by a question and a goal. Pro trick: don’t ask vaguely “what is this?” Give a role and a goal - “you are a car mechanic, explain to a newbie whether driving is dangerous.” Hidden trick: first ask the AI to describe WHAT it sees in the photo, and only then draw conclusions - this way you catch the recognition error before you believe the advice. A typical rookie mistake: attach a photo without any text at all and wait for the mind to be read. AI sees the pixels, but does not know WHY you showed the picture.