Voxel Artist Loop

Overview

Green tests do not mean good art. In the project that produced this skill, every artist agent passed 100% of its tests while shipping a wolf that read as "a human in a purple morphsuit" and wings that read as earmuffs. The render is the deliverable; the screenshot review IS the test. Spend the strongest model you have on the artist (the top tier produced a step-change over the next one down on the same prompts), and never commit art you haven't looked at yourself.

The Loop

1. Reference first. Give the agent an actual image file (agents that can read images will look at it) and name the bar: "match this pack". Order it to study the reference BEFORE reading code. No reference → agents anchor on the existing (bad) art.
2. Contract + fences. State the hard invariants of your engine (for example: limbs stay inside their columns, feet at y=0, palette-only colors, no module-level state) and exact file ownership. Shared files are append-only.
3. Mandatory look-iterate. The prompt must require: screenshot → look at the screenshot → compare to reference → iterate, with an expected round count ("expect 2–4 rounds"). Say explicitly: "a green test with an ugly character is a failure."
4. Report format demanded up front: files + net lines, per-item voxel/size budgets, iteration log (what changed each round), palette/shared-file additions, gate output lines, and at least 5 enumerated mismatches vs the reference per round.
5. Orchestrator reviews the pixels. Open every final render yourself before committing. Send delta feedback naming the exact defect and the fix direction ("fur must dominate — garment becomes a torso-only vest").

Prompt ingredients that mattered

• Resolution ceiling stated as fact: detail is capped by grid density; artists cannot stamp a monocle into a 7-voxel face. Fix the engine headroom before blaming artistry.
• Palette before modeling: all colors from a shared palette module; artists may append ramps, never edit existing exports.
• Perf budget as a number ("<20k voxels per figure", "60 ms mesh time"), not "keep it fast".
• Poses/builds sweep: require a robustness pass across every body build, pose and animation frame — it caught adornments floating mid-air in one pose that all default-pose screenshots missed.
• Gates before renders: a clearance check (no part intersects another across all clips) and a perf check must pass before any render is shown.
• Determinism: generators use hash-based noise, never Math.random, so re-running a generator gives the same model.

Observed failures this skill counters

Lessons from a full art audit

• Render-pipeline parity first. Judge art only in the renderer that ships. An unlit, fixed-shade, black-background stage made a purchased pro asset pack look as muddy as home-made boxes; a lit path lifted every model by about a point. Fix the stage before another art round: real lights (Lambert + hemisphere + a warm front-top key), a soft contact shadow, neutral tone mapping, a gradient backdrop.
• Procedural boxes plateau around 5/10. Seven rounds of parametric painter code never passed 5. Every charming voxel humanoid (Crossy Road, Kenney, Minecraft, commercial packs) is hand-authored geometry; code only picks parts, swaps palette and places them. Commission hand-made .vox parts, not more painter tweaks.
• Score against a fixed bar. 0–10 where a reference asset's store render = 7; re-score ALL models after each renderer or art change, in a table. "Looks better" without a number is not a review (see the judge-loop skill).
• Measured sculpt rules for a small voxel human: 10-wide face + 1-voxel hair shell; nose 1 voxel proud with its own shaded side; eyes 3×1 white/pupil/white with a 2-voxel bridge; 3×1 brows 2 rows up; 4-wide neck with 2–3 skin rows visible; garments 1 voxel proud with a 1-voxel dark gap; limbs detached at joints. Hair and headwear never intersect: headwear trims the hair under its band, and the full hair × headwear matrix is tested, not spot-checked.

Red flags — stop and review the pixels

• "Tests pass, committing" without opening a render
• Accepting a report's self-assessment of visual quality
• Reference described in words instead of an image file
• Artist prompt without an iteration requirement or round count