The 10% rule
Most thumbnail impressions happen in sidebars and phone feeds at tiny sizes. Judge every candidate zoomed out to roughly 10% — if the emotion and subject don't read instantly, it fails, no matter how good it looks full-screen. Prompt for what survives shrinking: 'extreme close-up face, exaggerated shocked expression, bold rim lighting, simple high-contrast background'.
One subject beats three. Every extra element halves the attention each one gets.
Leave room for text — then use almost none
If you overlay text, prompt 'clear empty space on the left third' and keep it to three words maximum in a heavy font. The title below the thumbnail already carries the details; thumbnail text is a punch, not a sentence. Generate the empty-space version and add text in your editor so you can A/B different wordings on one base image.
Batch-test like the big channels
Top channels test 3–5 thumbnails per video. With generation this fast, you can too: same prompt skeleton, varied emotion ('shocked' vs 'triumphant' vs 'skeptical'), varied background color. Upload alternates over the video's first days and keep the winner. This single habit compounds channel growth more than almost any production upgrade.