Why I built this.
Most AI critiques of images fail in a specific way. They describe the image in neutral, adjective-heavy prose — "warm lighting, shallow depth of field, moody atmosphere" — and call this criticism. It is not criticism. It is description.
A working colorist, asked to critique a still, does something different. They name a decision, locate it in the frame, say whether it worked, and if it did not work, say what it should have been. The language is specific, the judgment is decisive, and the critique is actionable. This is what I wanted to see if a prompt chain could produce.
How it works.
The app is built on Gemini 2.5 Pro through Google AI Studio. The prompt structure is three layers deep. The outer prompt establishes voice and critique format. The middle prompt walks through a six-axis evaluation (light, composition, color, focus, skin, narrative). The inner prompt forces the output into decisive sentences rather than neutral description.
I spent about six hours getting the voice right. The rest was pipeline plumbing.
What it does well, and what it doesn't.
It does well on staged images with clear intent. It catches lighting inconsistencies with surprising reliability, names them specifically, and explains what a fix would look like. It does moderately on composition. It does poorly on narrative — the "what is this frame saying" question is still largely beyond the model, which is unsurprising and is in fact the topic of several pieces on this site.
Use it as a second opinion, not a first one. If you have an image you're uncertain about, run it through. The critique is blunt. It is also cheap. If the critique annoys you, that is useful data about what you already believe.
Build notes.
Full build notes, prompt chain, and a short video walkthrough are available at the source link above. The prompt is MIT-licensed. Take it, fork it, rewrite the voice, make it yours. I'd love to see what other practitioners build from the same starting point.