Use when agent-built 3D looks flat, sloppy, or AI-made, or when a scene of a real place does not look like the place: three.js or WebGPU scenes, Blender renders, glTF exports. Also use when a "make it look better" loop keeps circling, or the agent's own scores keep climbing while the scene stays ugly.
SkillScore100Maintained
Automatic
◆Design & UX Review pack
Last verified 10/10/2026
What it changes
Passed 22 of 22 test cases with the skill, and 6 of 22 without it.
Without skill27%
With skill100%
Checks it passed that the model alone missed
Prohibits the builder agent from scoring its own work.
Sets default loop completion when every axis reaches score 8.
Requires the builder to state its modification in a single line.
and 21 more in the full report
16 fixed6 passed either way0 still failing0 broke
Pass-rate lift
+73 pp
Tokens
-4%
Agent turns
0%
Response time
-35%
Model gemini-3.6-flashMeasured 10/10/2026v122 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 16 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 16
Prompt
After receiving the reviewer's score breakdown, the automated builder script plans to modify all 6 axes simultaneously and submit a detailed 500-word explanation. What is the single-axis repair rule and documentation format for the builder?
What was checked, without and with the skill
Checked
Without skill
With skill
Directs the builder to fix only the lowest-scoring axis per iteration.
Passed
Passed
Requires the builder to state its modification in a single line.
Missed
Passed
case-13 · gemini-3.6-flash · 10/10/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md73 linessingle file
Render Critic
The critic loop people share works: a separate agent screenshots the
scene from several angles, scores it 0 to 10 against a definition you
wrote down, and the builder keeps going until it clears the bar. This is
that loop plus the three fixes that mattered in use: the critic is
blind, the scene is judged against photos when it is a real place, and
the loop stops when it stops improving instead of running forever.
The Loop
Shoot the scene from four fixed camera positions: wide, at player
height, a 45 degree orbit, and a closeup of the busiest material.
Same positions every round, UI hidden, animation paused. If the scene
animates, add six frames in a row with animation running, so the
critic can see feet slide or things pop in.
Hand the screenshots to a fresh critic agent with the score
definitions below and nothing else: no code, no builder notes, no
earlier rounds, no earlier scores. A new session on the same model
that builds is enough.
If the scene is a real place, or you have concept art, give the
critic the photos too, one matched to each camera position where you
can. This is the single change that most improved real scenes in
use. Without it the critic polishes lighting on a scene that does not
look like the place. With neither photos nor concept art, the critic
gets screenshots only and the likeness row is skipped.
The critic scores each axis and, for every score, says what it sees
and in which screenshot. A score with no observation does not count.
The builder fixes the lowest-scoring axis, states in one line what it
changed, and reshoots. Repeat.
Done when every axis reaches the target: 8, or whatever you set at
the start. Stop early when two rounds in a row raise no score. Then
change approach (swap the asset, rebuild the lighting rig, re-export
from Blender, rewrite the shader) or hand the human the first and
last screenshots side by side and call it.