Take the tool-using agent from the previous page and remove the tools. What is left is a persona, and a persona is not a property of the model, it is a paragraph of English sitting above the conversation. The box on the right is that paragraph, and it is editable. Rewrite it, send another message, and you are talking to something else entirely, with no retraining, no fine-tuning, and no new weights. That is the cheapest and most misunderstood lever in applied AI.
Everything that makes this character a character is in this text. Edit it and send another message.
the model is the same claude-sonnet-4-6 on every preset. only this textarea changes.
Each message you send is appended to an array of turns, and the whole array goes to the model on every request along with the contents of the system-prompt box, because the API itself is stateless and remembers nothing between calls. What feels like memory is simply the transcript being re-sent each time, which is why a long conversation costs more than a short one and why clearing the array is all it takes to produce genuine amnesia. The system prompt is not a special mechanism either: it is text placed in a privileged position, so the model conditions every reply on it without it appearing in the visible transcript. Swapping presets swaps only that text, and the personality change you see is the same model attending to different instructions. Compared with the tool agent on the previous page, this loop has no tool-call parsing and no iteration, which makes it the smaller half of what people mean by an agent: character without capability.