What counts as a prompt
In a chat interface, the prompt is usually the message a user submits. The application may combine it with conversation history, a system prompt, retrieved memories, and formatting before the model sees it. The visible message is therefore not always the model's complete input.
In text-to-image generation, a prompt describes desired subjects, composition, style, lighting, or other properties. Image and video systems may also accept reference media, masks, pose maps, or other forms of conditioning. Those inputs guide the result alongside the text rather than turning every control into prose.
How a model uses it
A language model's tokenizer converts prompt text into token IDs. A chat template can add role markers and control tokens, after which the model predicts a continuation within its context window. Image pipelines commonly pass tokenized text through a text encoder and use the resulting representation to condition a denoiser. The exact path varies by model family and application.
A prompt influences generation rather than querying a fixed record. The same prompt can produce different results when the model, seed, sampler, temperature, conversation history, or other settings change. Models also differ in vocabulary, training, and instruction-following behavior, so syntax that works well for one model may transfer poorly to another.
Where it appears in adult AI
Adult AI chat products expose prompts as messages, character instructions, scenario fields, or persona descriptions. Image and video tools expose a main prompt field and may pair it with a negative prompt, prompt weighting, model-specific trigger phrases, or visual controls. A hosted service may also modify or reject prompts through its content policy before inference begins.
A useful prompt states the intended subject or behavior clearly enough for the selected model and interface. It does not override missing model knowledge, guarantee anatomical or character consistency, or establish consent to use a real person's likeness.
Prompt versus adjacent terms
- System prompt. Higher-priority instructions supplied by a chat application to define behavior, rules, or persona. A user's message is still a prompt, but it occupies a different role.
- Chat template. The model-specific formatting that turns role-based messages into the token sequence expected by a chat model.
- Prompt weighting. Syntax or controls that adjust the relative emphasis of prompt elements in supported image interfaces.
- Negative prompt. Separate text conditioning used by some image pipelines to describe what should be discouraged.
Limits and common misconceptions
Longer is not automatically better. Extra clauses can compete for attention, exceed a context limit, or add concepts the model represents poorly. Prompt behavior is also implementation-specific: parentheses, weights, special tokens, and role labels only have the meaning assigned by the interface and model.
A prompt is not the same as training or fine-tuning. It conditions a particular inference request without permanently changing the model weights. Applications may retain submitted prompts for product operation, moderation, or model improvement, so users should check the service's published privacy and retention terms before entering sensitive information.
Technical context
These dossiers show how a visible prompt becomes tokens and conditioning in image pipelines or a formatted token sequence in chat stacks.