Render Critic

Render Critic

The builder does not grade its own work here. Asked to, it leans on the description of the work, and an open-ended "loop until 8/10" inflates the score while the scene stays ugly. This skill splits the roles and pins everything the grade could leak through: same shots, same rubric, worst axis wins, capped rounds, failure allowed.

When To Use

Any agent-built 3D output judged by eye, from a single glTF prop to a whole three.js level or Blender render. Run it after the scene loads and animates correctly. It scores looks. Gameplay and frame rate are out of scope.

Setup: Keeping The Critic Blind

The critic is a fresh agent with no conversation history. Its entire input is the rubric section below plus the screenshot files. It never sees code, the builder's notes, prior rounds, or prior scores.

Get that empty context however your harness allows: a subagent, a fresh exec, a new session, a clean chat with the rubric pasted in. The mechanism is the fresh start, never the tool name. The critic starts on the cheapest tier with vision as a budget default, not a tested quality floor; step it up the first time its receipts get vague or it misses a flaw you can see. Save the strong model for the builder.

Shot List (identical every round)

Fixed resolution (1280x800), a fixed lighting state and time of day, all UI hidden, so every receipt has to point at the scene and never at a loading tile or a button. Pause animation and reuse four saved camera poses so the stills compare round to round. In three.js with WebGLRenderer, capture the canvas in the same frame as a render call, or set preserveDrawingBuffer: true, or the shot comes back blank. In Blender, render headless with a pinned config (example: Cycles, 96 samples, a fixed seed, AgX view transform) so the same script gives the same render every round.

Animated scenes add a motion burst. Whatever moves in play (a run cycle, a chase, water, particles) hides its worst tells between frames: foot sliding, pop-in, LOD popping, shadow shimmer. With animation RUNNING, capture six frames at a fixed interval (about 150ms) from the gameplay camera, named round-N-burst-k.png, and hand them to the critic as an ordered strip. A static prop render skips this.

Method

1. Round 0: shoot the shot list and have the critic score the scene as-is. This baseline is the before/after receipt.
2. Builder fixes the lowest axis first. Before the next round it states, in one line, what changed. No stated change, round forfeited.
3. Reshoot the same shot list.
4. Improvement gate: take the file the last report cited for the failing axis, pair it with the new shot from the same camera, and hand a fresh critic the two unlabeled. It gets the axis name, that axis's anchor row, and the question: which one is better on this axis. Ask twice, order swapped. The new shot must win both; a split or a loss counts as no improvement, and the round is forfeited. For the motion axis, compare the old burst strip against the new one the same way.
5. A fresh critic instance scores the new round against the full rubric, with no history of prior rounds.
6. Repeat to pass or cap. Cap: 5 scored rounds. Forfeited rounds (builder side: no stated change, or a failed improvement gate) burn a cap slot; voided rounds (critic side, see below) do not. At the cap the critic issues PASS or FAIL-FINAL. FAIL-FINAL is a valid ending; the rubric is never renegotiated mid-run.

Rubric (each axis 0-10; anchors are binding)

Scoring rules:

• Overall = the minimum axis score, never the mean. One bad axis fails the round.
• PASS = every axis at 8 or above.
• Every score cites its file and a pixel-level receipt. No receipt, no score.
• Scores between anchors interpolate; when unsure, score lower.
• Absolute scores drift between critic instances. The improvement gate, never the score delta, decides whether a round moved.
• Motion + continuity is scored from the burst strip, not the stills. A static render has no motion; skip that axis and take the minimum over the axes that apply.

Calibrate The Critic

Prose anchors drift between critic instances. Pin one reference image per axis: a shot you judge a true 8 on that axis, handed to the critic beside the rubric as "this is an 8 for lighting." It costs one image per axis and gives the critic something to compare against instead of a sentence to interpret.

Signs The Loop Is Gamed

• the score rose and the builder stated no change
• a score row cites no file or no pixel receipt
• the builder proposes edits to the rubric or the shot list
• scores land at exactly 8 across the board in round 4 or 5

Any of these voids the round. Rerun it with a fresh critic; a voided round does not count against the cap, its rerun replaces it.

Critic Report (fixed format)

round 2 verdict: FAIL (overall 5; 2 rounds left)
| axis | score | file | receipt |
|---|---|---|---|
| lighting + tone mapping | 5 | round-2-wide.png | sky blown to white, foreground uniformly gray |
| materials | 6 | round-2-closeup.png | column albedo is one flat value, roughness identical to floor |
| silhouette | 8 | round-2-gameplay.png | columns, walls, roof slab separate cleanly |
lowest axis: lighting + tone mapping. fix that first.

Fix Map (builder starts here on a failing axis)

Stack Notes

• Unity, Unreal, or Godot: the loop and the rubric should transfer unchanged, untested here. Swap the capture mechanics and the tone mapping names for your engine; the notes below cover three.js and Blender, the stack this skill was built on.
• three.js: fix lighting and materials before reaching for the renderer. WebGPU buys compute and density headroom, not taste; a flat WebGL2 scene ports to WebGPU still flat.
• Blender to web: glTF with transforms applied, textures packed or baked, meters as units. Headless bpy with a pinned render config is the path this skill was written against and keeps renders repeatable; a live MCP plugin is untested for the shot list.
• Postprocessing: at most one pass (bloom or AO) until every axis clears 7. Effects come last; keep the receipt the critic cites visible until the axis passes.

Cost

One loop = builder rounds plus, per round, one cheap critic call on four stills (six more frames when the scene animates) and a two-image improvement gate asked twice. The expensive version (a strong-model critic, a dozen shots, no cap, scores argued in the main chat) multiplies spend. The cap is the budget; the gate is the signal.