Agnes Model Prompt Engineering

Core Principle

The essence of video prompting: clearly tell the model what you want to see, what''s happening, and what mood you want. Unlike text generation, video prompts must describe visual content, camera movement, and temporal change. The following 5 tips go from easy to advanced.

#1Scene First — describe the setting before the subject

The model needs spatial context first. Start with the environment (location, lighting, atmosphere), then the subject and action. A good prompt reads like the first line of a storyboard.

#2Action Specificity — use verbs, not adjectives

The model understands specific actions far better than abstract adjectives. Instead of ''dancing gracefully'', say ''spins on her toes, skirt flaring outward''. Every action should be something you could draw frame by frame.

#3Camera Movement — tell the model what the camera is doing

The Agnes model supports camera movement descriptions. Add camera instructions (drone shot, close-up, tracking shot) at the beginning or end of your prompt.

#4Temporal Sequence — use punctuation to describe time progression

For videos over 10 seconds, use periods or semicolons to separate events happening at different times. The model processes them in order.

#5Negative Prompts — tell the model what NOT to include

If the model repeatedly generates unwanted elements (text, watermarks, specific colors), use ''avoid:'' or ''negative:'' at the end of your prompt.

Common Mistakes

• Too short

  ''A video of a cat'' — the model lacks context, output is random. Include at minimum: setting, action, and atmosphere.

• Too abstract

  ''Show the eternity and sorrow of love'' — models can''t understand abstract concepts. Convert to concrete imagery: ''Two people embrace on a rainy train platform as the train slowly pulls away.''

• Mixed languages

  ''A cat running in the park'' mixed with other languages — mixing languages can cause semantic conflicts. Use all one language.