Skip to main content

How does the prompt system work?

C
Written by Chris

A prompt is the written description you provide to guide an AI generation. It tells the model what you want to create, such as the subject, setting, visual style, mood, or movement.
​

Blueberry provides access to several third-party AI video models. Each model interprets prompts differently, so the same description may produce different results depending on the model you select.


A clear prompt helps communicate your idea, but it does not guarantee that every instruction will appear exactly as described.

How does a model interpret my prompt?
​

When you submit a prompt, the selected AI model uses your description to generate a result based on its capabilities and how it has been trained.


The model may give more attention to some parts of your prompt than others. It may also interpret an ambiguous phrase differently from what you intended. For example, "a cinematic scene" leaves many decisions open, while "a wide shot of a quiet city street at sunset, with warm lighting and slow camera movement" gives the model more specific direction.
​

Adding relevant details can help. However, a longer prompt is not always more effective. Too many competing instructions can make it harder for the model to produce a consistent result
​
​What is Blueberry's role?
​
Blueberry is the platform through which you access and use the available models.
For generations powered by third-party models, the underlying provider's technology determines how your prompt is interpreted and what the model produces.


This means that Blueberry cannot directly control every creative decision made by a third-party model or guarantee that it will follow every instruction precisely.


We understand how disappointing it can be when a result does not match your description or is not usable for your intended purpose. A mismatch does not necessarily mean you wrote your prompt incorrectly; it can also reflect the selected

model's limitations or interpretation.
​
​Why doesn't the result always match my prompt?
​
AI generation involves variation. Even a detailed prompt can produce unexpected results, and using the same prompt again may lead to a different output.

Common reasons include

  • Ambiguous wording: The model may choose a different interpretation of a subject, action, or style.

  • Competing instructions: Several actions, visual styles, or camera movements may be difficult to combine.

  • Too many requirements: The model may capture the main idea while missing smaller details.

  • Model limitations: Some models may struggle with precise movement, spatial relationships, consistency, or complex

  • sequences.

  • Differences between providers: An instruction that works well with one model may be less effective with another.
    ​

Think of your prompt as guidance for the generation, rather than a set of commands that the model will always execute exactly.

How can I write a more effective prompt?

  1. Start with the main subject and action. Describe what should appear and what should happen. Make the most important part of your idea easy to identify. Example: "A golden retriever walking along a quiet beach."

  2. Add the setting and visual details. Include details that meaningfully shape the scene, such as the location, lighting, colors, and mood. Example: "A golden retriever walking along a quiet beach at sunrise, with soft golden light and gentle waves."

  3. For video, describe movement clearly. Specify the subject's movement and, when relevant, the camera's movement. Keep the action focused. Example: "A golden retriever walks slowly along the shoreline at sunrise. The camera tracks alongside the dog in a smooth, steady shot."

  4. Keep instructions consistent. Try to use descriptions that work together. For example, asking for both a stationary camera and an orbiting camera in the same shot may lead to an unpredictable result. If your idea includes several separate actions or scene changes, start with a simpler version and assess how the selected model handles it.

  5. Describe what you want to see. Use direct, concrete wording. Instead of "make it look better," describe the desired change, such as "soft natural lighting, a simple background, and a close-up view."

  6. Refine one element at a time. If the result misses your intention, simplify the prompt or adjust one part before trying again. This can help you identify which wording communicates your idea more clearly. Each new generation may still produce a different result.

Example: From a broad idea to a focused prompt

  • Broad prompt: "A dragon flying over a castle."

  • More focused prompt: "A red dragon glides above a medieval stone castle at sunset. Warm orange light illuminates its wings, with mist covering the mountains in the background."

  • Video-focused prompt: "A red dragon glides steadily above a medieval stone castle at sunset. The camera follows from the side in a smooth, wide tracking shot. Warm orange light and distant mountain mist create a dramatic fantasy atmosphere."

These details give the model clearer direction, but the final result will still depend on the selected model's capabilities and interpretation.

What should I do if the model repeatedly misses my instructions?
​

Start by simplifying your prompt and making the main subject and action explicit. If another suitable model is available, you can also try adapting the prompt for that model.
​

If the behavior continues, contact our support team with your exact prompt, the model you selected, and a description of how the result differed from your instructions. Include the generation reference or output if available.
​

These details help our team review the behavior and, where appropriate, raise it with the third-party provider. Reporting an issue helps us investigate, but it does not guarantee a specific outcome or a change to the model's behavior

Did this answer your question?