openai-quick

OpenAI Quick Image Generator

Generate a single image fast, without session overhead. Best for one-off images, prototypes, or exploration.

When to Use This

• Creating a single image with no plans to iterate
• Generating a single new image anchored to an existing reference (without starting a session)
• Exploring different concepts (generate several variations)
• When you don't need organized file management
• Quick reference images or placeholders

When NOT to Use This

• If you'll want a series of consistent variations → use the openai-sprite-series skill instead (sessions + images param)
• If you're building a multi-screen project → use the openai-ui-mockups skill instead
• If you want to iterate and refine → use continue_editing after the first generation

Steps

1. Ask the user what they want to generate (if not already described).

2. Check API status with get_configuration_status. If not configured, guide through setup:

   • Set OPENAI_API_KEY env variable (recommended), or
   • Run configure_openai_token with their key from platform.openai.com/api-keys

3. Choose parameters based on the request:

   • aspectRatio: 1:1 (square/social), 16:9 (wide/presentation), 9:16 (mobile/story), 3:2 (photo), 21:9 (cinematic)
   • model: gpt-image-2 (default, best quality) — switch to gpt-image-1.5 only if transparent background or inputFidelity control is needed
   • quality: low for quick exploration, medium for normal work, high for production output

4. Generate with generate_image. Write a descriptive prompt:

   • Include: subject, style, lighting, mood, composition, color palette
   • Avoid vague keywords — "a cozy coffee shop with warm amber lighting, steam rising from cups, watercolor style" not "coffee shop"

5. If anchoring to an existing reference — add referenceImages: ["/absolute/path/to/ref.png"] (up to 16 refs, all paths must be valid — any invalid path is rejected before the request is sent):

   • gpt-image-2 (default): always high-fidelity, suitable for style anchoring; does NOT support inputFidelity param — omit it
   • gpt-image-1.5: add inputFidelity: "high" when identity preservation matters (exact subject likeness); supports background: "transparent" if needed
   • SERIES with refs → skip this pattern and use sessions + send_creative_message instead (see openai-sprite-series or openai-workflows)

6. Visually verify with Read tool. If the result needs tweaking, use continue_editing with a correction prompt.

7. Report back with the image path and any notable details about what was generated.

Prompt Quality Guide

Vague ❌	Descriptive ✅
"a cat"	"A ginger tabby cat sitting in a sunlit window, watercolor illustration, warm tones"
"a logo"	"Minimalist tech startup logo, letter 'G', geometric, dark blue on white, vector style"
"space scene"	"Deep space nebula, purple and gold clouds, distant stars, photorealistic, cinematic lighting"

Universal Prompt Formula

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

Example: 2D pixel art wizard character [subject], centered, front-facing [composition], casting a fireball [action], plain white background [location], clean outlines, limited 16-color palette [style], 32x32 pixel grid [technical]

Iteration Strategy

• Wrong detail → continue_editing("change X to Y") — targeted correction
• Wrong composition → start fresh with generate_image — editing rarely fixes fundamental layout
• Style drifted → pass the original image as a reference in continue_editing
• Text errors (e.g. "SWIM" → "SVIM") → always verify text with Read tool, then correct with continue_editing

Speed Levers

When turnaround time matters more than final polish:

• Use quality: "low" first for draft exploration — escalate to "medium" or "high" only for finals
• jpeg renders faster than png when you don't need transparency — use png only when an alpha channel is required
• Square (1024×1024, the default) is the fastest shape — non-square sizes add a small latency overhead

(Source: OpenAI cookbook)

See the openai-prompts skill for domain-specific prompt structures (sprites, UI, icons, characters).