Simply what you type in. Everything the model produces is a continuation of it, which is why its wording matters more than most people expect. In most applications the prompt also includes instructions set by the developer that the user never sees, shaping how the model behaves.
Technical view
The full prompt typically comprises a system message defining role and constraints, any conversation history, retrieved documents, and the current user turn. Every part consumes context and influences output. A significant security concern is prompt injection, where instructions embedded in content the model reads, such as a web page or an uploaded document, are treated as directives. This is a live and largely unsolved problem for any system that reads untrusted input.
Related
Part of AI Terminology Index, section: Building and Using Models.