Fun
AI
Don't tell my spouse but she is getting some upgrades to her makeup station for Christmas.
She has a small area dedicated to makeup and hair. Its a desk with a single mirror on the wall. The idea is to add mirrors on each side of the current mirror, angled, that will help her see the sides and back when she is styling her hair.
Below I asked Nano Banana Pro to drop the chosen mirrors into the picture to give you an idea of what I have in mind.
Wait a minute though. How do I know that adding these mirrors in this space is going to actually solve the problem? I can tell that in the configuration above, they would not be effective for helping her to see the back of her head when styling her hair. Is there a way to tell if the mirror dimensions, wall position, available pivot angles will all work to actually allow her to see everything? I don't want to find out after I have purchased the mirrors and drilled the holes in the wall.
For these kind of the projects, I have a time-honored approach: yolo it. Normally, would I know, a priori, that the geometry is going work out such that the mirrors will reflect all the right stuff? No, I would not. Look, I studied light and optics once upon a time, I should be able to figure it out, I guess. Angle of incidence is equal to the angle of refraction, as I recall.
But its 2025, almost 2026. Why think, when you can prompt?
And blessed as I am, with access to 3 state of the art models at my fingertips, why just prompt, when you can bakeoff?
So here are, dear reader, the results of my Make Up Station Mirror Simulator Bakeoff of ChatGPT 5.2, Gemini 3 Pro Preview, and Opus 4.5
Preview
Building the Prompt
ChatGPT 5.2 got the honor of helping me construct a prompt I could then use with each of the three models to build the simulation. I gave it some detail about the dimensions of the area, the dimensions of the mirrors, my objective. Honestly, this all happened in the morning before my first coffee and I didn't overthink it:
Help me create a good prompt that I will be able to use to instruct the model to build me a simulation for some changes I am making to my wife's makeup station. There is already a chair, a small desk and a mirror with lighting on each side. I want to add a mirror on each side so that she is able to see the back of her head when styling her hair. What I need to simulate is the size, position and tilt angle of the mirrors in order to enable her to see all angles of her head. The relevant dimensions are: makeup desktop 20" x 29", current mirror 23" wide x 29" tall. The desk is in a corner with a short wall on one side and the other side open. The short wall extends 11" from the wall behind. The distance from the current mirror to the short wall is 7 inches.
The resulting prompt is here. I am not sure if I read it carefully. Like I said, pre-coffee.
Time to one-shot this with each of the models.
The initial ChatGPT 5.2 result looks polished and seems extensive in the set of things you can control, but I had some difficulty figuring out how it was supposed to work. Gave up on it before I got any value back from it. Fail
The initial Opus 4.5 result seems so promising. Fairly intuitive to control the placement of the mirrors. Still can't make anything work regardless of geometry. Fail
This initial Gemini 3 Pro Preview result almost kinda works. A few key issues. We'll see if we can get those with an updated prompt. Round 1 Winner 🏆
Rather than iterating interactively which each model, which I suspect would have got me to a positive outcome much faster, I decided to update the prompt and try to one shot it again, and rerank winners.
Round 2
Prompt update instructions:
Update the prompt with these additional instructions:
- in the top down view, the wall with the current mirror is at the bottom and the short wall is located to the left
- The head can be modelled to be an exquisitely beautiful but statistically-speaking typical human head for all dimensions
- For default mirror, retrieve dimensions and any other details from https://www.wayfair.ca/decor-pillows/pdp/winston-porter-corrente-pivot-n-view-squared-cornered-rectangle-mirror-for-window-bathroom-vanity-c009207780.html?piid=1993632662
- Try to default the initial position and angles of mirrors to something that works
- keep the key controls to make adjustments to mirror position, size and tilt on screen so I don't have to scroll to make those adjustments
The resulting prompt is here. Still not reading that crap. Let's one shot it again.
ChatGPT 5.2 result was pretty nice. A little too many configurable parameters, and the eye seems to be able to point off is some pretty unnatural directions. Overall its not quite helping me figure out if the mirrors I am getting are going to work but maybe it would if I got serious about working with all the controls. In tight race , Round 2 Winner 🏆.
Opus 4.5 result was a bit more approachable but also had issues with the gaze looking off in extremely unnatural directions.
Gemini 3 Pro Preview result is approachable but like the other models has done an insufficient job of determining all the pertinent characteristics of a human head.
Round 3
Prompt update instructions:
Further adjustments:
- the position and dimensions of walls are not modifiable for this simulation
- the size, shape of the head and position of eyes are not modifiable for this simulation
- to simplify we are only interested in simulating on one plane, can omit controls related to height and vertical tilt of mirrors
- as a further constraint the mirrors must be attached to the back wall. Check the dimensions of the specified mirror product to determine offset and pivot point
- the side wall should be considered a constraint on available pivot angles for closest mirror (it cannot pivot through the wall
- my wife is able to turn her head, in order to bring the back of her head into the field of view.
- she is also gifted with the ability to gaze into mirrors other than the centre mirror if it helps see something
The resulting prompt is here. More coffee now but quick skim of that only.
ChatGPT 5.2 something must have gone wrong in the prompt, one of the mirrors is embedded in wall.
Opus 4.5 eye positioning is weird.
Gemini 3 Pro Preview has the subject seated with the back of their head towards the mirror, which is unconventional, and ultimately useless. So like the others its a Fail but grudgingly will declare this the winner of this round 🏆.
Round 4
Prompt update instructions:
Refinements
- Please consider my wife's head beautiful but also opaque
- My wife is able to lean in towards the back wall to position her beautiful head a distance of 10" away from the back wall
- The mirrors are to be mounted on the back wall
- The mirrors should match the dimensions and details of the specified mirror, and any controls to modify mirror dimensions should be hidden in an advanced setting section
- The mirrors are mounted on the back wall, and the product has brackets that offset the mirror from the wall, with a pivot point in the center of the mirror. The reflective surfaces should be oriented towards the beautiful subject and controls should allow for the yaw to be adjusted from -180 to +180 degrees
The resulting prompt is here. Let's one shot this again with this hopefully improved prompt.
ChatGPT 5.2 Promising looking set of controls, but seems like there is a bug, the actual mirrors and head are not rendering for me.
Opus 4.5 Positioning of the mirrors starts in a bad place and can't be made to work.
Gemini 3 Pro Preview seems promising but ray tracing doesn't seem to work.
None of these are good, not declaring any winner this round.
Declaring Prompt Bankruptcy
After the last set of results I was not certain that things were getting better. Have had a full coffee by now, let's start over and try to make a better, more focused prompt.
Make a prompt that I can use with different models to create a simulation that will help me determine if some changes I am considering making to my wife's makeup station will have the desired outcome. I want to add two new mirrors, one on each side of the existing mirror, in order to allow her to see the back of her head when styling hair at the makeup station. The key question is whether the mirrors I have in mind, when mounted in the current space, and pivoting at angles that work within the space constraints, will allow my wife to see the back of her head.
My wife's head is exquisite and beautiful but statistically typical in terms of human head size, position of eyes, and opacity. She is able to lean forward in the space to position the front of her lovely noggin about 10" away from the central mirror. She is able to turn her head and cast her precious gaze to look into any of the mirrors available.
The space is pictured in attached image. My intention is to place a mirror, mounted on the back wall, on each side of the central mirror. The mirror I am planning to buy is https://www.wayfair.ca/decor-pillows/pdp/winston-porter-corrente-pivot-n-view-squared-cornered-rectangle-mirror-for-window-bathroom-vanity-c009207780.html?piid=1993632662. Retrieve mirror dimensions from this website. Note this mirror has a mount that allows it to pivot about a vertical axis through its center. The mount is perpendicular to the wall, with the edge of the mirror offset 4" from the wall and the pivot point of the mirror offset 15" from the wall.
The mirrors will be mounted on the back wall, with reflective surfaces oriented in towards where my wife is seated. The mirrors are not able to pass through the walls, so the mounting position and pivot range for the mirrors is constrained by wall dimensions. There is a desk in a corner that defines the workspace. The relevant dimensions are as follows: desktop: 20" deep x 29" wide, current mirror: 23" wide, perpendicular to the back wall, a side wall protrudes 11" perpendicular to the back wall. The distance between the current central mirror and the side wall is 7"
The simulation should allow me to modify the yaw of the mirrors, the position of my wife's head and gaze, and indicate whether the back of her head will be visible. It should try to position the head and mirrors in a way so that the starting point is a configuration with good posterior head visibility.
Favor a longer prompt that includes more of the detail I have provided here to allow models to reason out the most useful solution.
My desired outcome is a simple simulation that I can load in my browser that will show a 2D top down view that indicates to me the effectiveness of the proposed mirror installation for full head viewing. I don't want to buy these mirrors and install them only to find out later the optical angles are suboptimal.
This diatribe resulted in producing this prompt. Still not reading it.
ChatGPT 5.2 comes close but doesn't really work. The offset of the mirrors from the wall is wrong and cannot be adjusted. So it a fail, but probably the winner of this round 🏆
Opus 4.5 Nice but the ray tracing seems to not work.
Gemini 3 Pro Preview One of the mirrors can't be yawed the right way. The eyes are looking away from the central mirror which makes it a struggle to get any ray tracing to work.
Final Round
Adding some refinements to the previous prompt
Also specify:
- The mirror has an overall width of 16" and the width of the mirror itself is 14".
- The pivot axis is offset from the wall by half the width of the mirror itself (7") plus the 2" offset to accommodate the mounting bracket for a total of 9"
- The pivot allows a full 360 degree pivot of the mirror
- One surface of the mirror is reflective and this is positioned to point in towards the makeup station
Which resulted in this prompt.
ChatGPT 5.2 effort is not working. Maybe the eyes are on the wrong side of the head? Not sure what all is going on, but ray tracing doesn't seem to work.
Opus 4.5 Almost working. The problem seems to be that the back of the head is modeled to be where the front of the head ought to be.
Gemini 3 Pro Preview. Boom. There it is. This works. It answers my questions. The mirrors can be positioned realistically and as the head turns it shows the portions of the noggin that are visible via a reflection in the mirror. 🏆 Mission accomplished and Winner Overall🏆
Conclusion
I don't usually work with one shot prompts. Instead I tell the machine what I want, with kind of medium-level precision, and ask it to make a plan. I review the plan, it fills in the gaps in my specification, I suggest some changes. Together we get to a well-defined solution, iteratively. It doesn't have to be specified all up front.
Nonetheless, I do find the one shot prompt to be an interesting test of model performance. Almost always, when you want a AI to do something of value to you, the prompt is going to be inadequately specified, and will sometimes contain conflicting guidance. The model that feels the most intelligent is the going to be the one that somehow cuts through the imprecision and contradictions and delivers what you really needed to solve the problem, in spite of what you said or what you failed to say.
In this particular test, not-very-scientific, but real world to me, Gemini 3 Pro Preview came out ahead as the winner.
