openai-characters

OpenAI Game Character Generator

Full workflow for creating consistent game characters across multiple poses, sessions, and quality settings. Combines sprite workflow, pixel art rules, and prompt structures.

Model choice comes first

• Isolated sprites / transparent assets → model: "gpt-image-1.5" + background: "transparent". The default model (gpt-image-2) REJECTS transparent backgrounds — switch models before generating.
• Full scenes, references, max quality → default gpt-image-2.

---

Step 1: Define the Style Bible

Before generating anything, establish these rules. Paste as a prefix in every prompt for this character.

Resolution: [32x32 sprites, 16x16 tiles]
Palette: [PICO-8 / 16 colors]
Outlines: [1px black]
Shading: [2-tone cel, cool-shifted shadows, warm highlights]
Proportions: [2.5 heads tall]
Background: [transparent / white]

Name every visual detail of the character: colors, armor pieces, markings, accessories, distinctive features. If it's important, name it — the model won't invent consistent details.

---

Step 2: Start a Session

start_creative_session({
  model: "gpt-image-1.5",          // use gpt-image-1.5 if transparent background needed
  aspectRatio: "1:1",
  description: "[Character name] sprites",
  outputDirectory: "/path/to/project/assets/characters/[name]"
})

Use a fixed outputDirectory inside your project — this is your style anchor folder across sessions.

---

Step 3: Generate the Canonical Base Image

Use send_creative_message with a fully detailed first prompt:

[Style bible prefix] [character type] in idle stance, front-facing,
[every visual detail: colors, armor, markings, accessories],
[background], [technical: pixel size, palette, outline, shading, proportions]

Example: 2D pixel art, 32x32 sprite, PICO-8 16-color palette, 1px black outline, 2-tone cel shading, cool-shifted shadows, transparent background — knight in idle stance, front-facing, blue tabard with gold lion emblem, silver plate armor with horizontal visor slats, blue feathered plume on helmet, brown leather boots, 2.5 heads tall

Do NOT pass images for this first generation.

Visually verify with Read tool. This image becomes the canonical style anchor for the entire character.

---

Step 4: Generate Additional Poses

For every pose after the first, pass the previous image as reference:

send_creative_message({
  prompt: "Same [character], [pose] — maintain identical colors, armor design, and proportions",
  images: ["/path/to/previous-pose.png"]
})

Suggested pose order for chaining (each references the previous): Idle → Walk → Run → Attack → Jump → Hurt → Death

• Keep prompts short after pose 1 — visual details propagate via the image reference
• Verify each pose with Read tool before moving on
• If a pose is wrong, fix it with continue_editing before generating the next

Preserve-list at every step. Alongside images[], restate what must not change in EVERY chain prompt — "Same knight: same helmet, plume, tabard, gold lion emblem, proportions, palette. Change only the pose to ." Drift compounds when later steps rely on "same as before". For multi-ref steps, address refs by index: "Image 1: base identity. Image 2: previous pose. Keep Image 1's design; continue the motion from Image 2." (Source: OpenAI cookbook; complements the validated images[] chaining — see finding-gpt-image-workflows.)

(validated on GPT Image 2026-06-05: single-ref and two-ref chains produced production-level consistency on gpt-image-2 — see [[finding-gpt-image-workflows]])

---

Cross-Session Consistency (Returning Later)

Sessions expire after 30 minutes. When resuming in a new Claude session:

1. Pass the canonical base image (idle pose) as images[] in your first call
2. Paste the style bible text alongside it for maximum fidelity
3. The MCP has no memory between restarts — the canonical image IS the continuity

send_creative_message({
  prompt: "[Style bible] Same knight as reference, now in [new pose]",
  images: ["/path/to/project/assets/characters/knight/canonical-idle.png"]
})

Save the canonical image path — write it in the project CLAUDE.md or README so it's always findable.

---

Resolution Guide

Size	Use Case	Detail Level
16×16	Lo-fi, mobile, game jam	Silhouette only
32×32	Sweet spot for most games	Facial expressions, armor, weapons
64×64	Detailed indie games	Fabric folds, individual fingers

Generate in 32×32 first. Scale to other sizes by passing the base image:

send_creative_message({
  prompt: "[Style bible at 16x16] Same character, same pose, scaled to 16x16 — simplify to silhouette and key colors",
  images: ["/path/to/32x32-idle.png"]
})

---

Color Palette Reference

Palette	Colors	Best For
PICO-8	16	Retro platformers, game jams — the default
Sweetie 16	16	Mobile games, friendly/soft aesthetic
Endesga 32	32	Fantasy RPGs with varied environments
Apollo	16	Sci-fi shooters, cool/tech aesthetic
Zughy 32	32	Horror, dungeon crawlers
GameBoy	4	Demakes, nostalgia, maximum constraint

Hue shifting: Don't darken with black — shift shadows toward blue/purple. Don't lighten with white — shift highlights toward yellow/orange.

---

Model Choice

• GPT Image 2 (gpt-image-2) — default for full scenes, references, and max quality. Preferred for consistency-critical sprite series. No transparent backgrounds. Arbitrary sizes up to 3840px (ratio ≤ 3:1), always high-fidelity. Up to 16 input refs.
• GPT Image 1.5 (gpt-image-1.5) — use when transparent backgrounds are required (isolated sprites). Preset sizes only (auto, 1024×1024, 1536×1024, 1024×1536). Supports inputFidelity param. Caution: showed proportion drift on single-ref character edit in 2026-06-05 validation (chibi shift B2 vs B1) — not recommended as the primary model for pose-consistency chains.

(validated on GPT Image 2026-06-05: gpt-image-2 produced production-level cross-pose consistency; gpt-image-1.5 showed proportion drift on ref-edit — see [[finding-gpt-image-workflows]])

---

Post-Processing (Industry Standard)

AI generates the base → manual refinement → final export:

1. Generate 3–5 variations of the base pose
2. Pick best silhouette/geometry
3. Open in Aseprite (paid) or Piskel (free)
4. Apply palette normalization (replace colors with locked palette)
5. Fix outlines, pixel noise
6. Export as PNG sprite sheet or individual frames