Generate Image with Gemini

Generate an AI image using the generate_image_gemini MCP tool. Before calling the tool, rewrite the user's prompt into a detailed, well-structured image description so the model produces a high-quality result.

Workflow

1. Understand what the user wants. If the request is vague ("make me an image"), ask one focused question to capture the subject. Otherwise proceed.

1. Optimize the prompt. Rewrite the user's description into a detailed image prompt following the structure below. Do this silently — don't show the user a wall of technical prompt engineering unless they ask.

1. Choose sensible defaults. Unless the user specified otherwise:

• quality: fast (Nano Banana, ~1K output) — cheap and quick, good enough for most requests. Only move up when the user signals they care about fidelity.
• aspect_ratio: 1:1 for icons/avatars/logos, 16:9 for landscapes/banners, 9:16 for portraits/phone wallpapers, 4:3 for photos
• format: png (lossless, safest default)

1. Call the tool. Use mcp__mcp-media-toolkit__generate_image_gemini with the optimized prompt. Only bump quality above fast when the user explicitly asks for higher fidelity, mentions production use, or asks for photorealistic/print-quality work — higher presets are much slower.

1. Show the result. The tool returns a thumbnail preview and the local file path. Report both to the user, plus confirm what aspect ratio and quality preset were used.

1. Offer upload. After generation, ask if they want a public URL. If yes, call mcp__mcp-media-toolkit__upload_image_s3 with the returned file path. Skip this step if the user already said they only want the local file.

Prompt Optimization

Short casual prompts produce generic output. A good image prompt describes subject, context, and style with enough specificity that two people reading it would imagine roughly the same image.

Checklist

Before sending a prompt to the model, make sure it covers these dimensions (skip any that don't apply — e.g. a logo doesn't need lighting):

• Subject: What is the main thing? Be specific about appearance, age, posture, expression.
• Context / setting: Where is it? What's around it? Foreground vs background.
• Composition / framing: Close-up, wide shot, overhead, eye-level, rule of thirds.
• Lighting: Natural, studio, dramatic, golden hour, soft, backlit, neon.
• Color palette: Warm tones, muted, vibrant, monochrome, specific hues.
• Style: Photorealistic, illustration, watercolor, cinematic, 3D render, minimalist flat design.
• Technical quality: High detail, sharp focus, 4k, depth of field.

Keep the final prompt to 1-3 dense sentences. Don't write paragraphs — the model responds better to information-dense phrasing than to long narratives.

Examples

Example 1 — vague request

• User says: "a cat on a book"
• Optimized: "A fluffy orange tabby cat curled up on top of an open vintage hardcover book, soft natural window light falling from the left, warm cozy tones, shallow depth of field, photorealistic, high detail."

Example 2 — product shot

• User says: "an image for my coffee app's landing page"
• Optimized (after asking what kind of vibe): "A minimalist overhead shot of a single white ceramic cup filled with latte art on a clean marble countertop, soft diffused morning light, muted warm color palette, lots of negative space on the right for text, cinematic, editorial photography style."

Example 3 — abstract / logo

• User says: "generate a logo concept for a dev tool called 'Pulse'"
• Optimized: "A minimalist modern logo for a developer tool named 'Pulse', a stylized heartbeat waveform forming the letter P, flat vector design, electric blue on white background, clean geometric lines, technology branding."

When NOT to over-optimize

If the user gives a very specific prompt already ("photorealistic portrait of a 40-year-old blacksmith with a leather apron, forge in the background, Rembrandt lighting"), don't add fluff — they know what they want. Your job in that case is just to pass it through and maybe tighten the wording. Only expand prompts that are genuinely underspecified.

Quality Presets

• fast (default) — Nano Banana (gemini-2.5-flash-image) at 1K. Cheap, quick, good enough for most drafts, sketches, and casual requests. Start here unless the user signals otherwise.
• balanced — Nano Banana 2 (gemini-3.1-flash-image-preview) at 2K. Better detail, slower. Use when the user cares about quality — "this is for a blog post", "for my landing page", etc.
• quality — Nano Banana Pro (gemini-3-pro-image-preview) at 4K. Maximum fidelity but significantly slower (can take several minutes). Reserve for when the user explicitly asks for photorealistic output, print-quality assets, or mentions maximum quality. Warn them about the wait.

Errors

If the tool returns an error:

• "Gemini API error" — likely a bad or missing GEMINI_API_KEY, or quota exhaustion. Tell the user to check https://aistudio.google.com/apikey.
• "returned no image data" — the model refused the prompt. Rephrase to avoid ambiguity or content policy edges.
• "File not found" — the output directory isn't writable. Suggest passing output_dir with an absolute path the user controls.