openai-prompts

OpenAI Prompting Guide

Write effective prompts for GPT Image generation. The core principle: describe the scene narratively, don't just list keywords.

Universal Prompt Structure

[Subject] + [Composition] + [Action/State] + [Location/Context] + [Style] + [Technical specs]

Bad: knight pixel art game Good: 2D pixel art knight character in idle stance, front-facing, blue tabard with gold lion emblem, silver plate armor, white background, 16x16 pixel grid, game sprite sheet style

For per-domain prompt structures (sprites, environments, characters, UI, icons, web sections), see references/domain-structures.md.

---

The Identity Guide Pattern (validated 2026-05-12)

For series work — sprite sets, evolution-chain atlases, mockup screens, item packs — pin a byte-identical identity guide block at the top of every call in the series. The variant prompt then describes only what differs.

Five-line structure:

Identity: [IP name, what the world is, who plays]
Form: [3D/2D, geometry style, surface character, proportions]
Camera: [angle, lens, key/fill lighting setup]
Palette: [Name #HEX, Name #HEX, ...]
Materials: [textures, finishes, special-state effects]

Why it works: the model treats the guide as canonical context for what world this is in. The variant prompt only has to answer "what specific item this time?". Cross-call consistency comes from the guide being literally the same text every time, not from passing image references.

The Identity line is load-bearing. Give it a specific fictional project name — "Hollowfest Manor," "Crystal Caverns," "Tideglass Cove," whatever yours is. Real or invented; both work. Don't settle for category descriptors ("a Halloween merge-2 game") when an IP name is available. The model's prior for fictional-IP-cohesion activates when there's a name to anchor on — empirically a stronger consistency anchor than abstract world descriptors alone. Adapted from Dori Adar's original framing of the pattern.

When to use:

• Multiple thematic asset variants in the same world (3D items, evolution chains, UI screens with shared brand)
• Cross-modality projects where the same IP needs to render across 3D items AND flat-vector UI

When NOT to use:

• Single-subject repetition across poses → use the Hybrid Workflow with image refs instead (see openai-workflows skill)
• Unrelated one-off images → the guide is overhead with no payoff

Validated across a 27-image A/B/C experiment + 6-deliverable production pass.

⚠️ Provisional: Identity Guide validation was run on gemini-creative; re-validation on GPT Image pending.

---

Prompt Anatomy (asset-series template)

For multi-image work, structure prompts in this labeled order — a skimmable template beats clever syntax. The Identity line comes first (load-bearing anchor; see above), then:

Identity: <IP/project name> — <asset family>
Scene: <background/context, or 'plain white background'>
Subject: <the thing, pose/state>
Details: <materials, palette, lighting>
Constraints: <what must NOT appear / must stay fixed>

Full guidance: references/openai-prompting-essentials.md

---

Text in Images

GPT Image handles text but demands precision:

• Put literal text in quotes or ALL CAPS; demand "verbatim, no extra characters"
• Tricky or invented words (brand names): spell them out letter-by-letter in the prompt
• Specify font style, size, color, and placement as hard constraints
• Use quality: "medium" or "high" for small or dense text — "low" quality garbles it
• Verify with Read tool — text errors are common (e.g., "SWIM" → "SVIM")

Full guidance: references/openai-prompting-essentials.md

---

Photorealism

• Include the literal word "photorealistic" to engage photo mode; alternatives: "real photograph", "taken on a real camera"
• Use camera/lens language: "35mm film photograph, 50mm lens, eye level, shallow depth of field, subtle film grain"
• Avoid "studio polish" phrasing — it triggers over-smoothing (Source: OpenAI cookbook)

Full guidance: references/openai-prompting-essentials.md

---

Prompting Rules

Do:

• Use narrative descriptions ("a cozy coffee shop with warm amber lighting" not "coffee shop warm")
• Name every distinctive visual detail to preserve across iterations
• Specify style explicitly (photorealistic, watercolor, pixel art, flat vector, 3D render)
• Describe lighting and mood ("dramatic rim lighting", "soft diffused daylight")
• Use positive framing ("clean white background" not "no clutter")
• For series work: pin a byte-identical Identity Guide block at the top of every call (see above)
• For dense-label or atlas-shaped prompts: append the negative-instruction suffix verbatim — "Do not render color swatches, hex codes, or palette legends anywhere in the image." Validated at 5/6 (production pass) + 2/2 (#14b A/B) ≈ 88% suppression.

Avoid:

• Pure keyword lists ("cat space neon cyberpunk")
• Vague quality requests ("make it better", "improve the details")
• Assuming the model will maintain details that weren't described — if it's important, name it
• Re-describing the same style with different words across calls in a series — empirically causes consistency drift. Pin the guide verbatim instead.
• "For game asset extraction" / "for reference" framings inside the prompt — observed correlating with the palette-legend rendering bug; phrase in product terms instead ("for a Halloween merge-2 game", not "for asset extraction")

---

Isolation Suffix for Asset Extraction

When generating individual items destined to be cut out and used as standalone sprites/icons, append this hard-rules block verbatim:

Hard rules — must be followed exactly:
- Single object centered in frame, fully visible with breathing room from edges
- Pure white #FFFFFF uniform background, no gradient or vignette
- NO cast shadow, NO drop shadow, NO ground shadow — the item floats on the flat background
- NO color swatches, NO hex codes, NO palette legends, NO color charts anywhere
- NO UI elements, NO text, NO labels, NO decorations beyond what is described

Production-validated across 33 generations: ~100% shadow compliance, ~95% swatch suppression. The output is directly consumable by a chroma-key extraction pipeline. For white-ish subjects (ghosts, white skulls), fall back to semantic segmentation (rembg). See the openai-game-assets skill § Single-Isolated-Subject Extraction Pattern.

⚠️ Provisional: extraction validation rates inherited from gemini-creative; re-validation on GPT Image pending.

---

Iteration Strategy

1. First pass: Generate with a full detailed prompt
2. If detail is wrong: Use continue_editing with a targeted correction ("change the helmet visor to horizontal slats")
3. If overall composition is wrong: Start fresh with generate_image — editing rarely fixes fundamental composition issues
4. If style drifted: Pass the original image + style description, not just the correction

---

Additional Resources

Reference Files

• references/domain-structures.md — Per-domain prompt structures: game sprites, game environments, character concepts, UI/mobile mockups, icon sets, web sections.
• references/openai-prompting-essentials.md — Distilled OpenAI cookbook best practices (anatomy, preserve-lists, ref indexing, text accuracy, photorealism, quality strategy, iteration).