Prompt Tips

The quality of your AI-generated output depends heavily on how you write your prompt. This guide covers prompt writing techniques and the built-in optimization feature.

Prompt Optimization

modelBridge includes a one-click prompt optimizer. Click the sparkle button below the prompt field, and an AI rewrites your description into a model-optimized prompt — without requiring prompt engineering expertise.

• Cost: approximately $0.01 per use (shown in hover tooltip)
• How it works: your prompt is sent to an LLM that rewrites it with better structure, specificity, and model-appropriate language
• The button is disabled when the prompt field is empty and re-enables on first keystroke
• Visual feedback: shimmer animation during optimization, green flash on success, red flash on error

The optimized prompt replaces your original text. If you prefer the original, undo with Cmd+Z.

General Prompt Writing Tips

Be Specific About Motion

For video models, describe the movement you want — not just the scene.

Instead of…	Try…
”A city at night"	"Slow aerial dolly forward over a city skyline at night, neon lights reflecting in wet streets, light rain"
"Ocean waves"	"Close-up of waves crashing on rocks in slow motion, spray catching golden sunset light, cinematic"
"A person walking"	"A woman in a red coat walking through a rainy street, camera tracking left, shallow depth of field”

Describe Camera Work

AI models respond well to cinematographic language:

• Camera motion — “slow pan left”, “dolly forward”, “crane up”, “static wide shot”, “handheld tracking”
• Framing — “close-up”, “medium shot”, “wide establishing shot”, “over-the-shoulder”
• Focus — “shallow depth of field”, “rack focus from foreground to background”

Specify Lighting and Mood

• Lighting — “golden hour”, “harsh midday sun”, “neon-lit”, “overcast soft light”, “backlit silhouette”
• Mood — “peaceful”, “tense”, “dreamlike”, “gritty”, “whimsical”
• Style — “cinematic”, “documentary”, “vintage film grain”, “clean corporate”

For Image-to-Video

When animating a still image, focus on motion rather than describing the image:

• Good: “Gentle wind blowing through hair, subtle head turn to the left, soft smile”
• Less effective: “A portrait of a woman with brown hair” (the model already sees the image)

For Text-to-Speech

TTS models respond to punctuation and formatting cues:

• Use commas and periods to control pacing
• Use ellipses (…) for dramatic pauses
• Write phonetically for unusual pronunciations
• For ElevenLabs v3, use emotion tags like [excited], [whispers], [laughs]

For Sound Effects

Describe sounds with specificity:

• Good: “Thunder rumbling in the distance, followed by heavy rain on a tin roof, occasional wind gusts”
• Less effective: “Storm sounds”

Prompt Persistence

Your prompt text carries over when you switch models. Write a prompt, switch from Kling v2.5 to Kling v3 Pro, and the text is already there. Model-specific settings (duration, aspect ratio) reset to the new model’s defaults, but your prompt stays.

Iterating on Results

If the first result is not what you wanted:

1. Adjust the prompt — add more specific details about what was wrong
2. Try the optimizer — it may restructure your prompt more effectively
3. Switch models — different models have different strengths and interpret prompts differently
4. Adjust parameters — sometimes changing duration, resolution, or aspect ratio produces better results than changing the prompt