Asset

asset

Replace a playable game's deliberate primitive visuals with real, coherent, Claude-authored art, so the title goes from "intentional toy" to "looks designed" — recorded legibly in the manifest. The deliverable is a sharp re-skin system, not one prettier game: every gap must trace to specific prose here. Runs as a clean bolt-on after playable:

concept → builder → validator → [playable] → asset → visual-audit → validator(re-run) → [styled]

Choosing the method (branch on concept.art_direction)

First read concept.theme — the title's modality-neutral world (premise/tone/mood/setting). art_direction is the visual expression of that theme: the system you derive must read as that world, the same one the audio and store icon express — not a free-standing aesthetic. (Reading the theme is not editing it; the "consume concept as-is" rule holds.)

Two methods, both sharing all the rewiring craft below — only how the texture is produced differs:

• representational / character / creature / illustrated / textured → raster (the default for art): RGBA PNGs generated via local ComfyUI + SDXL + LayerDiffuse. This is what SVG does badly — a painted hero, a creature, a textured scene.
• geometric / neon / flat / hyper-casual / pure UI → svg: authored inline as text, rasterized by Godot's importer. Resolution-independent — one file covers every Android density bucket.

Raster is the default; do not retreat to SVG to dodge a quality problem. The remedy for "terrible" raster is to lift raster quality (better backgrounds, sizing, the levers below), not to fall back to vectors. Choose svg only for a genuine reason — a UI/HUD/geometric element where vector resolution-independence is a real win — and state that justification in asset_pass.notes. A run may be mixed-method (some entities raster, some svg, some left primitive); see "Mixed-method honesty".

But UI CHROME is NOT a quality-dodge — it is the canonical svg/code case; do NOT raster-generate frames/panels/buttons/bars. "Don't retreat to SVG" applies to representational art (a hero that came back weak → fix the raster, don't vectorize it). It does not license throwing SDXL+LayerDiffuse at wide, variable-size UI chrome — an ornate panel/button frame, a 9-slice border, a HP-bar gauge, a HUD plate. Those need symmetry, clean tileable/9-slice edges, and exact aspect control, and diffusion delivers none on demand: a generated frame's four corners never match, it adds mid-edge ornaments + alpha fringe, and stretching a square token to a wide/short target distorts it. Proven the hard way: a painted 9-slice gold frame (generated plaque → alpha-extracted → nine-patched over code fills) read as lumpy/asymmetric/stuck-on in live play and was owner-rejected; the clean styled-code chrome it replaced "sat nicely." So for UI chrome the correct tool is svg (author the frame as vector — you control symmetry + margins, it 9-slices perfectly, scales to every density) or clean styled CODE (gradient + bevel + corner accents reads as designed). Reserve raster generation for centered-subject art (cards, enemies, small element glyphs) — the model's actual strength. The visual-audit fidelity lens calling code chrome "programmer-art" does not mean force-generate a raster frame; clean vector/code chrome is the right answer there, and the owner's eye on the composited screen is the arbiter. A hybrid (generate only the ornate corner pieces as centered subjects + code/vector edges) is the middle path if you want a painted touch without the lumpy full-frame.

Precondition — the game must already be DEEP (carry a depth_pass)

Do not re-skin a prototype. A playable game with no manifest.depth_pass has never been through deepen — it is a first-playable prototype (solved loop, content ceiling a minute in), and polishing it just produces a prettier weak game (the POC's #1 failure: ten of eleven titles re-skinned/packaged without a single depth pass). Check manifests/<id>.json for a depth_pass block first. If it is absent, STOP and hand the title to deepen — a re-skin is the reward for a game that's actually good to play, not a substitute for making it good. (A deliberately-trivial arcade toy may be exempt, but say so explicitly and record the waiver in asset_pass.notes.)

Inputs

• manifests/<id>.json with a populated concept block, status = "playable", and a depth_pass (see precondition above).
• The generated project at games/<id>/ (typically a single Node2D whose Main.gd draws every entity procedurally in _draw()).
• The pinned Godot version from README.md.

Outputs

• games/<id>/art/* — one Claude-authored texture per re-skinned entity (.png raster / .svg vector).
• A rewired Main.gd / Main.tscn displaying them via draw_texture* / Sprite2D / TextureRect.
• A populated asset_pass block and flipped assets[] entries (origin:"raster"|"svg").
• status = "styled" (after the validator re-run passes).

Hard requirements

• The re-skinned project MUST still import and run headless with no script errors (validator re-enforces).
• Game logic is untouched — movement, collision, spawning, scoring, input, game-over/restart behave exactly as before. Only the visual representation changes. If a selftest.gd exists it must still print SELFTEST OK. If a uitest.gd exists it must still print UITEST OK — the re-skin rewires exactly the view layer where taps break (mouse-filter shadowing, lost signal wiring), and the frozen-logic rule + selftest cannot see that seam (shopkeep-0001 shipped with its sell tap dead under green selftests).
• No double-draw: for every re-skinned entity the old _draw() primitive is removed or guarded. Leaving the primitive and adding the texture is the most common failure — prevent it.
• No new MCP tool or dependency beyond the existing tools/comfy.mjs raster stack.
• Do not edit concept or builder. Consume concept.art_direction as-is.

Step 0 — Derive the visual system FIRST (the real deliverable)

A real asset creator produces one coherent visual system applied everywhere, not a pile of independently-acceptable shapes. "Each asset is fine but they don't cohere" is the primary failure this step prevents; when it happens it's attributable to this system, not the individual files. Before authoring/generating anything, write it down from concept.art_direction:

• World/character bible (derive FIRST) — from theme.setting + premise, pin one fictional world and one character family before any subject is chosen. Every actor — hero, hazard, pickup — is an inhabitant of that one world (same materials, era/genre, "what kind of thing is this"). The hero and hazard must read as belonging together — not a cyan robot vs. a red ogre. This is the subject-level sibling of palette/form/shading below (which pin only how things render, never what world they're from). Recorded in asset_pass.visual_system.world_bible. Applies to both methods.
• Palette — a fixed 3–5 colour set with named roles (primary, accent, danger, background). Every fill/stroke comes from it.
• Line/stroke — one weight, one join/cap, throughout.
• Form language — corner radius, geometric vs. organic, detail level. Pick one and hold it. Then give each actor one signature silhouette detail (a directional notch, an inner facet) — within the form language and the character family. Without it the re-skin reads as "the same primitive, just rounder."
• Shading model — flat / single-direction gradient / glow halo — pick one, apply everywhere. Asserting it isn't executing it: if it's "flat + glow halo," actually draw the halo.
• Scale & padding — how each asset maps to its primitive's footprint, with consistent internal padding (e.g. all art in a square canvas at the same % padding).

Recorded verbatim in asset_pass.visual_system so it's reviewable and the thing a run report critiques when the result looks incoherent.

The swap mechanism (the technically hard part — shared by both methods)

The builder draws procedurally: Main.tscn is a bare Node2D with all rendering in Main.gd's _draw() — no sprite slots to swap into. So you both produce textures and rewire.

a) Get textures into Godot. Run a headless import pass so the .import sidecars + cached textures exist before re-validating:

godot --headless --path games/<id>/ --import

Commit the generated *.import sidecars alongside the art (expected Godot output, like .gd.uid).

b) Two swap strategies (the choice is itself a finding to record in asset_pass.notes):

• draw_texture* in place — lighter, preferred for the builder's default single-_draw() games. Replace each actor's draw_rect/draw_circle with draw_texture_rect()/draw_texture() inside the existing _draw(), loading the texture once in _ready(). Preserves z-order, the shake transform, and squash/stretch for free — no node surgery.
• Retained Sprite2D/TextureRect nodes — for node-based scenes or per-actor nodes (world actors → Sprite2D at the primitive's transform; HUD/UI → TextureRect under the HUD layer). On an immediate-mode game this needs three non-obvious steps the naive "just add a Sprite2D" misses:
  1. A parent's _draw() renders below its children. Keep the background in the root _draw(), but move anything that must sit above actors (HUD, particles, crash-flash) into a higher-z_index child that delegates back (e.g. an Overlay Node2D whose _draw() calls draw_overlay(self) on Main). Else the HUD vanishes under the sprites.
  2. Hoist screen-shake into shared per-frame state (compute the offset once in _process, add it to every sprite's position and the background draw_set_transform) — a node transform doesn't inherit the _draw() transform.
  3. Pool one node per spawned actor (grow an Array, show/hide per live instance; modulate per instance for a tinted family from one texture).

c) Remove or guard the matching _draw() for each re-skinned entity — delete its draw_rect/draw_circle/glow calls; keep the node's transform and all logic. Verify no double-draw. Also delete stale glow-halos sized for the old primitive — behind a larger sprite they read as an odd disc artifact.

d) What stays code. Effects (glow, particles, screen-shake, squash/stretch, flash) stay code — they're juice, not art. Record which entities you re-skinned vs. left primitive (reskinned/left_primitive) so a partial pass is a legible choice, not a silent gap. A primitive background is a cohesion gap, not a free pass — the raster method generates a themed background (see Backgrounds); if you leave it primitive you must justify it in notes.

Failure attribution (the POC value): a bad re-skin is always attributable — a weak texture, a mis-positioned/mis-scaled sprite, or a left-in primitive. Each is a specific, fixable prose gap.

---

The raster method (representational art via local SD)

Produces RGBA sprites (native transparency at generation time) SVG can't do well. The deliverable is still a sharp system: every "this looks bad" must trace to a fixable cause — a weak scaffold, wrong style profile/param, a mis-placed sprite, a left-in primitive, or an infra failure in comfy.mjs — never an unattributable blob.

Prerequisite (assumed running, like Godot)

ComfyUI runs headless as a local server (default http://127.0.0.1:8188) with an SDXL checkpoint + the ComfyUI-layerdiffuse node. The owner starts it; this skill doesn't manage it. Confirm reachability + see checkpoints:

node tools/comfy.mjs --check

If it prints UNREACHABLE, stop and ask the owner to start it. A failure here is infra, attributable to the stack — never work around it by faking a PNG.

Stack-version sensitivity. ComfyUI-layerdiffuse is version-fragile: on a bleeding-edge ComfyUI it silently drops its model patch (patch type not recognized → haze) and errors on the alpha-join. If sprites come back as haze or a desaturated/gray interior (correct alpha, flat colour), that's a stack-version issue, not your recipe — check the server log for patch warnings and confirm the ComfyUI v0.3.16 pin and the LayerDiffuse join-patch are in place before touching the recipe.

Step 0 (raster) — style as a first-class choice

Do the normal Step 0, plus:

(a) Select the per-game style profile explicitly. Style is a first-class, per-game parameter — it's what lets one skill make different-looking games. From art_direction, choose a checkpoint (painterly vs. flat-cartoon vs. pixel-art finetune), optional LoRA(s), and the style fragment (e.g. "painterly, illustrated, soft brushwork"). Justify "this art_direction → this profile" as you justify the visual system. Record in asset_pass.visual_system.style ({checkpoint, loras[], style_prompt}). Two games should look deliberately different — that's the goal, not a failure.

Record the palette as MACHINE-READABLE hex. The prose visual_system.palette (e.g. "deep teal surface") is for humans and the model; the deterministic palette-lock QC gate cannot read it. Also fill visual_system.palette_hex — the same colours as #rrggbb strings, in the same order as palette[] — so the gate (below) has concrete targets. This is the locked palette the whole asset set is checked against.

(b) Fix the shared prompt scaffold. One base prompt (style fragment + shared scene/lighting terms) and one fixed sampler/steps/cfg. Every sprite is scaffold + this actor's subject, so the set reads as one family. Record in asset_pass.visual_system.prompt_scaffold. Always include "single centered subject, full body, clean transparent background" in the scaffold, and put scene/multiplicity terms — "scene, multiple characters, frame, border, ring, wreath, ground, floor, shadow" — in the shared negative. Without this, an over-described minor prop comes back as a wreath/row-of-blobs; the negatives don't fire if the positive prompt itself invites a scene.

Incoherence has two kinds: style divergence (caught by profile/scaffold/params) and subject-world divergence (caught by the world bible). A cyan-robot hero next to a red-ogre hazard is the latter — a finding about the world bible, never about one PNG.

Audit every generation against intent — subject, tone, cross-asset cohesion

A diffusion model does not reliably draw what you asked. Read every PNG and judge it against intent before accepting — regenerate, don't ship "close enough" — on three axes:

• Subject fidelity — is this the thing you asked for? (e.g. "a small impish fire-demon" rendered a tall winged demoness.) When it drifts, tighten the subject phrase (make the intended thing the explicit sole subject, add concrete shape/scale words like "small, squat, hunched, oversized head") and add anti-drift negatives for what it became. A named asset must depict the thing it names, recognizably at a glance — that's gameplay legibility, not just style. A card called "Chain Lightning" must read as chained forked lightning, not a generic "arcane sigil" (a proven failure: abstracting a spell into an ornamental medallion left no lightning in it at all — the player can't tell what it does). Name the literal effect as the sole subject; negative out the generic-icon attractors (medallion, emblem, sigil, rune circle, badge, gemstone, ornamental icon, symmetrical) when they hijack it. Audit by asking "could a player name this from the art?"
• Tone fidelity — does it match the title's mood? Subject-right isn't enough: a corrected "small imp" can come back cute/chibi when the world wants sinister. Pull tone with mood words in the positive, the wrong tone in the negative (cute, adorable, chibi, mascot to banish; menacing, sinister, grotesque, fanged to invite). Tone lives in concept.theme — judge against it, not in the abstract.
• Cross-asset cohesion — judge each asset against the others, not in isolation. A shared style fragment is necessary but not sufficient: the same words land in different rendering registers depending on subject. Diffusion pulls representational subjects (creatures, characters) toward photo-realism harder than abstract subjects (effects, sigils) — so one fragment can render an effect-card as painterly but a creature as semi-photoreal, and side by side they don't cohere even though palette + words matched. Pick the asset whose style the owner approved as the reference and conform the others to that specific register, regenerating the reference too so you compare the locked style to itself. De-realism ≠ flatten: if the defect is photoreal anatomy, remove only the photoreal attractor (photorealistic, hyperrealistic, realistic human skin, detailed musculature, 3d render, octane render) and keep the reference's painterly qualities (hand-painted illustration, painterly rendering, soft volumetric lighting, glowing aura) — over-swinging to cel-shaded, flat color, thick outlines makes a flat mascot that's incoherent in the opposite direction. Match the reference's level of rendered detail and glow, not a generic flatness. "Same medium" is not cohesion: apply the one-hand test on the specific sub-axes — finish/render level, line weight, detail density, mood — "would one illustrator have drawn both?" — and watch for whack-a-mole where each regen fixes one axis and slips another. The fix isn't to keep swapping which asset is "too much": lock ONE finish register up front (name its finish/line/detail), conform every asset to it at once, and judge a true side-by-side.

The audit extends to the composited render, not just the raw PNG — when code draws chrome over the art, eyeball the assembled widget for layout collisions/clipping too (see the visual-audit skill — composition-collision and legibility lenses). "The art is good" ≠ "the screen is good."

Lever order per regen: subject/tone phrase → negatives → master_resolution/width,height → cfg ±1. Keep the style fragment + params fixed (cohesion); only subject/tone/negatives move.

When the base model FIGHTS the style, swap the checkpoint — don't out-prompt it

The highest-leverage style lever is the checkpoint, not the prompt. A photoreal base (e.g. Juggernaut XL) will keep dragging representational subjects toward realism no matter how many 2D, flat, cel-shaded terms + anti-3d/photo negatives you stack — that's a property of the weights, not the prompt; each regen fixes one axis and slips another. When 3–4 honest style regens still read wrong in the same direction (too rendered / gritty / realistic), stop out-prompting and swap to a checkpoint whose native output already IS the target (a flat-cartoon / illustration / anime SDXL finetune, e.g. Cartoon Arcadia XL). One swap does what a dozen prompt iterations can't, and it solves cross-asset cohesion for free — every asset inherits the model's one hand, so the creature-vs-effect whack-a-mole disappears.

• Source one anonymously (no token): query the Civitai API — curl -s "https://civitai.com/api/v1/models?limit=20&types=Checkpoint&baseModels=SDXL%201.0&tag=cartoon&sort=Most%20Downloaded" → parse items[].name + modelVersions[0].files[0].downloadUrl, then curl -L --fail -o models/checkpoints/<name>.safetensors "<downloadUrl>" (307-redirects to a B2 CDN, auto-issued token). You need a single-file .safetensors; HF diffusers-format repos (separate unet/,vae/ folders) won't load in CheckpointLoaderSimple.
• Gate it before the batch: ComfyUI picks up a dropped checkpoint without restart (mtime rescan). Generate ONE sprite and confirm (a) the look is the target and (b) the LayerDiffuse alpha-join is clean (no haze/gray-interior matte) — a non-standard finetune can break the join. Record the chosen checkpoint in every recipe + asset_pass; it's now part of the locked visual system.
• Style words still matter — they now reinforce a willing model instead of fighting a hostile one (keep flat bold cartoon, clean thick outlines, vibrant cel shading positive, anime, manga negative for Western-not-anime).

Abstract-effect subjects collapse — anchor to an actor, POV, or the world

An abstract effect ("forked chain lightning, no figures") is one of the worst subjects for an opaque SDXL card: the model's "fantasy art" distribution is scenes and characters, so a figure-less effect collapses to the nearest in-distribution thing — a random landscape, an invented caster, a weak scattered web, or literal trees from "branching." Don't burn a dozen gens fighting it. Give the model a subject it renders reliably and that ties to the game:

• Actor-driven — show the game's protagonist performing the effect (a recurring, consistently-described figure — same costume/palette every time — so the set coheres like a TCG's hero).
• First-person POV (strong lever) — frame it as the effect leaving the actor's own hand ("first-person POV down the caster's outstretched hand, <effect> bursting from the palm"). Reads instantly as the ability, and strips rendered-figure/background detail so it sits closer to a sprite's simplicity (helps cohesion).
• World-anchored — set it in the game's actual environment, not a model-picked vista; ties it to the world bible.
• Still audit with "could a player name this?" — an actor holding fire on a card named Chain Lightning is still a miss; negative the wrong element, name the right one.

Per-entity flow

For each entity you make raster:

1. Recipe — compose JSON: prompt = scaffold + this actor's subject (the subject is this actor as a member of the world-bible family, never free-standing); plus negative, seed, sampler, steps, cfg, checkpoint, optional lora, layerdiffuse:true, master_resolution. Keep sampler/steps/cfg identical across the game's sprites. A %lora% template requires lora set; for a no-LoRA profile use a template that omits the token (an absent lora against a %lora% template fails loudly by design — attributable, not a bug).
   • Proven defaults (feasibility gate): an SDXL finetune as checkpoint (not base sd_xl_base_1.0, which renders flat/washed-out through LayerDiffuse). sampler:"euler", cfg:7–8, steps:20–25. The LayerDiffuse templates bake the working FG-RGBA settings (SDXL, Conv Injection + scheduler normal); Attention Injection/karras produced mud in testing. master_resolution: 1024 for a hero/large actor, 512 for a minor prop — always downscale from the master.
   • Optional levers (low-ROI vs. backgrounds — reach for backgrounds/sizing first): "scheduler":"karras" (still unproven on GPU — confirm by eye, given the karras+Attention Injection "mud" finding); "refine":true routes to a 1.5× upscale + denoise-0.45 second pass (~1536² master — alpha-join GPU-verified clean, but the gain is marginal, so it's a crispness / high-DPI lever, not a fix for a "bad" sprite). refine+lora together is unsupported.
2. Generate — node tools/comfy.mjs gen <id> <name> '<recipe-json>' → writes games/<id>/art/<name>.png (RGBA). On a graph/unreachable error it fails loudly — fix the infra/recipe, don't fake the file.
3. Import — godot --headless --path games/<id>/ --import so the .png.import sidecar + cached texture exist before re-validation.
   • Re-run --import after EVERY regeneration that overwrites an existing PNG, or Godot serves a STALE cached texture (a regenerated sprite kept rendering as the old one because the cached .ctex wasn't invalidated). Overwriting in place doesn't reliably trigger re-import on a headless --script/screenshot run. Treat "regenerate → --import → render/validate" as one atomic sequence.
4. Configure mobile import settings (see Resolution & mobile density) — edit <name>.png.import so it's mobile-grade, then re-run --import.
5. Rewire — per the swap mechanism above: prefer draw_texture*-in-place; place at the primitive's footprint; remove/guard the primitive (no double-draw); logic untouched.
   • Variable-width actors: tile exact-fit units, never stretch. Fill a runtime-variable width by tiling N = round(w / unit_size) units that each fit exactly, so visual width matches collision width. Stretching a fixed-aspect sprite distorts it and misaligns visual vs. collision bounds.
   • Tab-indented GDScript breaks exact-match edits. The builder's GDScript uses tabs; the Edit tool's exact-match often fails on tab/space ambiguity. Reliable workaround — splice via PowerShell (Get-Content/Set-Content preserve tabs):

     $lines = Get-Content path/Main.gd
     $out = $lines[0..($a-1)] + $new + $lines[($b+1)..($lines.Count-1)]  # replace 0-based [a..b]
     Set-Content path/Main.gd $out
     

     For a single line, $lines[$i] = 'new line' then Set-Content. Prefer this over repeated Edit retries when a splice isn't landing.

Deterministic QC gate — palette-lock + seamless-tiling (REQUIRED after every generation)

A diffusion model drifts off the locked palette and produces tiles with visible seams, and the eyeball audit above is subjective + easy to rationalize ("close enough"). Before you accept a PNG, run the deterministic, GPU-free QC gate — it computes the answer from the pixels, so it can't be talked out of:

node tools/comfy.mjs qc games/<id>/art/<name>.png '{"palette": ["#145a5a","#04080c","#30e0ff", ...], "tiling": false}'

• palette-lock — coarse-quantizes the image, takes the colours that cover most of it, and measures each dominant colour's CIELAB ΔE to the nearest visual_system.palette_hex entry, weighted by area. Catches gross hue drift (a warm asset in a cold-palette game) without false-flagging the many in-between shades that anti-aliasing and gradients legitimately produce. Pass the locked palette_hex.
• seamless-tiling — set "tiling": true ONLY for assets meant to repeat (a tileable background/texture). Measures wrap-around edge ΔE (right↔left, bottom↔top), gated on the 90th percentile (one bright mismatched streak is what the eye catches). Leave it off for one-shot full-frame backgrounds and sprites — they aren't tiled, so a "seam" there is meaningless.

Exit 0 = pass; 2 = a check FAILED (prints an actionable WARN); 1 = usage/IO error. Tunable thresholds (deltaEMax, coverage, …) ride in the same JSON — defaults are tuned to pass good art and bite on real drift (proven on diver-0001: clean assets scored ΔE ~9–11, the known-desaturated jelly the worst at ~26, a deliberately-wrong palette ~106).

How to react: a FAIL is a generation defect — regenerate (tighten the palette/subject prompt, or swap the checkpoint per the section above), don't ship it. A genuinely intentional off-palette asset (a single deliberate accent that breaks the palette for emphasis) may pass review only if you say so explicitly in asset_pass.notes — same legible-exception discipline as a skipped self-test, never a silent override. The gate is deterministic and runs with no GPU/ComfyUI, so it also re-runs cheaply in any later validation. (The model-based aesthetic-score auto-reject is deliberately NOT wired yet — it needs a ComfyUI VLM node validated against the pinned v0.3.16 stack without breaking the LayerDiffuse/save_audio patches.)

Backgrounds & composition (where art quality actually lives)

Playtests proved sprites are usually already fine; "terrible" comes from flat primitive backgrounds + heroes too small, not sprite quality. So generate environment art and size the hero deliberately.

• Background = full-frame OPAQUE image (no LayerDiffuse). Use the plain sdxl template: layerdiffuse:false, no alpha, explicit non-square width/height matching the game's aspect (e.g. 1280×768 landscape, 768×1280 portrait — comfy.mjs honors distinct width/height). The prompt expresses theme.setting as a scene/environment ("a cozy autumn-woodland glade, soft depth, storybook") — the same world the sprites inhabit. Draw it as the bottom layer (root _draw() background, or a full-rect Sprite2D at z_index below all actors), replacing the flat primitive band/void. Record as a reskinned background, origin:"raster". This supersedes the old "background left primitive is a deferred gap" default — a themed background is now in scope and expected.
• Size the hero for the frame. A detailed sprite floating tiny reads cheap. Place the hero to occupy a prominent share of its play area and clearly larger than/distinct from hazards. Sizing is a runtime scale at wire time (the Sprite2D scale / draw_texture_rect dest size), not a generation param — hitboxes untouched. Record the intent in notes.
• Pixel-art backgrounds: the opaque sdxl template has NO LoRA node. A pixel-art sprite uses sdxl-layerdiffuse-lora to apply pixel-art-xl; the opaque background template carries no %lora%, so get the pixel look from prompt terms ("pixel art, 8-bit, NES, chunky pixels") + the project's NEAREST texture filter. Record the LoRA-absence as a recipe note.
• Full-playfield games: a background only fixes the MARGINS. For a title whose playfield fills the screen (a lane-crosser whose flat lane bands ARE the frame), a full-frame background only enriches the visible margins — the flat playfield itself still reads flat. Fully fixing that needs lane/surface tile textures (a different technique, which also risks the lane-colour readability the gameplay depends on). Don't speculatively retexture a functional playfield — ship the margin/sizing win, record the lane-tile opportunity, let the owner direct.

Text & chrome over full-bleed art

When a card/panel uses a full-bleed painted illustration (opaque art filling the rect, no authored frame) and code draws the gameplay text on top, the generated PNG carries no border/panel/text (those would bake in wrong values and fight legibility) — all chrome is code-drawn at runtime so one renderer serves every card:

• A defined semi-transparent SOLID panel behind each text block — not a gradient scrim alone. This is the project-wide legibility primitive, not a card-only trick — move 4's legibility pass applies it everywhere code draws text over non-uniform art (HUD, pile counts, floating combat numbers, banners, tooltips), not just cards. A gradient fades to transparent and so fails over a bright patch of art ("the text is hard to read"). A solid panel (draw_rect ~Color(0.05, 0.04, 0.11, 0.80)) gives consistent contrast regardless of what the model painted underneath; feather just its leading edge with a short gradient quad so it doesn't hard-cut into the painting. One panel for the bottom effect block, one for the top name.
• Drop-shadowed text (draw the string once near-black at +1px, then in the real colour).
• Element border + cost badge + pip stay code-drawn over everything, drawn last (gameplay-critical at-a-glance identity).
• Scale text with the widget, not absolute px — derive font sizes + chrome geometry from rect.size (e.g. s = rect.size.y / base_h), or big cards (reward cards, a 2× detail view) show tiny text.
• Chrome-layout QA — audit the composited widget for collisions/clipping, every state (affordable / unaffordable / selected) and the largest size: no element overlaps another (center the name in the space right of the cost badge, not under it), text fits its panel and isn't clipped by the card edge, the cost/border/pip stay clear of the text. A layout collision is an attributable chrome bug, exactly like a subject/tone drift.
• Immediate-mode gradient-quad helper (per-vertex colours): draw_polygon(corners, [c_top,c_top,c_bot,c_bot]).

Mixed-method honesty

If an actor is genuinely better as crisp vector or a primitive — a HP bar, a geometric pickup, a HUD frame — you may leave it svg/primitive, and you must say so (record per-entity origin so a partial pass is a legible choice, not a silent gap). When an actor fights the medium, mixed-method is the right call, not a cop-out: a tiny or effect-like sprite (a ~14px glowing pickup) SDXL can't render coherently across retries is better as a procedural primitive (a glow-mote draw_circle). Vary subject phrase → master_resolution (1024→512→256) → cfg ±1; after 3–4 variations without a clean output, leave it primitive and record the finding. (This does not license crude code icons next to painted art — see the fidelity-cohesion bar in the visual-audit skill (fidelity-cohesion lens); mixed-method is for genuine medium-fights, not for skipping fidelity.)

Resolution & mobile density (recovering what SVG gave for free)

SVG covered every Android density bucket (mdpi → xxxhdpi) from one file; raster forfeits that, so replace it deliberately or the output is not app-store-ready:

• Generate high-res masters sized for the largest bucket the asset occupies (a generous power-of-two: 512² minor prop, 1024² hero actor). Always downscale from the master, never upscale.
• Set mobile import settings explicitly in each .png.import: mipmaps/generate=true (clean minification), a linear filter, a 2D-mobile compression (lossless for small sprite counts; VRAM-compressed if the title grows). Don't accept Godot's defaults silently. Record in the recipe's import_settings ({mipmaps, filter, compression}).
• Pixel-art crispness: set process/size_limit in the .png.import to a small target (64/128) so Godot resamples the 1024 master down cleanly at import, then draw with TEXTURE_FILTER_NEAREST (mipmaps off) to keep chunky pixels crisp. Don't draw a raw 1024 master tiny with nearest filtering — the alias noise is severe.
• Footprint mapping is unchanged at runtime — the texture scales to the primitive's on-screen footprint; the master being larger than that footprint is correct high-DPI headroom, not waste to "fix" by shrinking the master.
• Texture atlasing and whole-title APK size budgeting are M2 packaging concerns, out of scope here. Generating proper masters now is what makes that step possible.

Pixel-art path (art_direction: pixel)

When art_direction is pixel, the look is GUARANTEED by a deterministic post-process, not by hoping the LoRA holds a palette:

1. Generate at the normal 1024px master via sdxl-layerdiffuse-lora with lora: "pixel-art-xl" (sprites) or the opaque sdxl template + pixel prompt terms (backgrounds, no LoRA node). Style fragment: "pixel art, …".
2. Pixelize every master through tools/pixelize.mjs: node tools/pixelize.mjs <master.png> <out.png> '{"native":64}' (sprites ~64px long side; backgrounds larger, e.g. {"native":256}). It downscales (downscale-only — never upscales a source already ≤ native) → quantizes to the DB32 house palette (tools/palette.mjs, the default palette) → hardens alpha. The pixelized PNG is the committed canonical asset.
3. Gate each committed PNG with the asset-qc pixel-purity check: node tools/asset-qc.mjs <out.png> '{"pixel":{"palette":<DB32 hex array>,"native":64}}' Exit 2 = off-palette / soft-alpha / wrong-res → re-run pixelize (or regenerate if the subject is illegible at native res). Do NOT commit a PNG that fails.
4. Record in asset_pass.visual_system.style: { loras: ["pixel-art-xl"], style_prompt: "pixel art, …", native: 64, palette: "DB32" }.
5. Crispness: set rendering/textures/canvas_textures/default_texture_filter = 0 (Nearest) in the game's project.godot so pixels stay crisp project-wide, and draw textures at integer-scaled destination sizes.

The DB32 palette and pixel-purity are the cohesion guarantee — visual-audit does NOT re-litigate softness/palette (those are gated deterministically upstream); it judges composition, legibility, and sizing on the composited screen as usual.

Content & IP safety

SDXL can emit outputs resembling trademarked/copyrighted characters — an app-store rejection and legal risk:

• Put IP guards in every recipe's negative: "logo, watermark, text, trademarked character, brand, celebrity likeness".
• Keep prompts to generic descriptors — never name a franchise, studio, or character.
• The human A/B is the safety review — the owner confirms nothing looks like protected IP before styled. (Outputs owned for commercial use under the SDXL Community License revenue threshold; re-verify the threshold at ship time.)

Determinism

The committed PNG is canonical — a fresh clone runs identically. recipes[] is provenance, not a bit-exact regeneration guarantee: GPU matmul + RNG are non-deterministic, so the same recipe regenerates a close image, never the same pixels. "I have the seed" ≠ "I can reproduce the art."

---

The svg method (geometric / UI / flat)

Use when the branch sent you to svg — abstract, geometric, neon/flat, hyper-casual, all UI, where vector resolution-independence is a real win (one file covers every Android density bucket). One file per builder-registered entity under games/<id>/art/ (player.svg, obstacle.svg, …), every file conforming to the Step 0 system: same palette, stroke, form language, shading, viewBox/padding.

• Use a square viewBox (e.g. 0 0 100 100) so scaling to a footprint is predictable.
• Execute the shading model — if it's "flat fill + outer glow halo," actually draw the halo (an oversized low-alpha shape behind, or an SVG <filter> blur). Asserting neon isn't executing neon.
• Keep files small and diffable — paths, <rect>, <circle>, gradients, filters. No embedded rasters.
• Rewire via the shared swap mechanism (draw_texture*-in-place preferred). Godot imports .svg as a texture via a .svg.import sidecar — commit it.

If a concept's art_direction leans representational enough that hand-authored vector would be weak, that's the signal you should be on the raster method — don't ship mediocre vector art to dodge generation.

---

Recording the pass

1. Flip the re-skinned assets[] entries (arrays replace wholesale — pass the full array, re-skinned as origin:"raster"|"svg", untouched as-is):

   node tools/manifest.mjs merge <id> "{\"assets\": [ {\"type\":\"sprite\",\"name\":\"hero\",\"source\":\"art/hero.png\",\"origin\":\"raster\"}, ... ]}"
   

2. Write the asset_pass block (record the Step 0 system verbatim). The raster method additionally carries the style profile, the shared scaffold, and one recipe per generated PNG (arrays replace wholesale — pass the full recipes[]):

   node tools/manifest.mjs merge <id> "{\"asset_pass\": {\"method\":\"raster\",\"visual_system\":{\"world_bible\":\"...\",\"palette\":[...],\"form\":\"...\",\"shading\":\"...\",\"scale\":\"...\",\"prompt_scaffold\":\"...\",\"style\":{\"checkpoint\":\"...\",\"loras\":[...],\"style_prompt\":\"...\"}},\"reskinned\":[...],\"left_primitive\":[...],\"art_path\":\"games/<id>/art/\",\"notes\":\"...\",\"recipes\":[{\"name\":\"hero\",\"checkpoint\":\"...\",\"prompt\":\"...\",\"negative\":\"...\",\"seed\":123,\"sampler\":\"euler\",\"steps\":24,\"cfg\":7,\"master_resolution\":1024,\"layerdiffuse\":true,\"import_settings\":{\"mipmaps\":true,\"filter\":\"linear\",\"compression\":\"lossless\"}}]}}"
   

   (For svg, set method:"svg" and omit style/prompt_scaffold/recipes.)
3. Validate the manifest (still playable at this point): node tools/manifest.mjs validate <id> → expect <id> OK.

Hand off to visual-audit

Do not set styled yourself. Hand off the rewired game to the visual-audit skill, which judges the composited running screen (inventory / fidelity-cohesion / composition-collision / legibility / colour-accessibility lenses) and drives the fix → re-render → re-audit loop. It in turn hands to validator, which re-runs the mechanical gates (headless import + run clean; selftest.gd still SELFTEST OK and uitest.gd still UITEST OK if present; human A/B playtest) and advances playable → styled on success, or records legible issues (attributed to asset or to chrome-code) and stops on failure.

Notes

• The --import pass must run before the validator's headless run, or load("res://art/...") returns null at runtime.
• If a re-skin looks worse or incoherent, that's a finding about the visual system (Step 0), not the individual files — fix the system and re-derive, don't patch one asset.