13 System prompts and parameter tuning

The chapter explains how to move a chatbot beyond a generic assistant by shaping its behavior with system prompts and tuning generation parameters. A system prompt is introduced as the hidden instruction that comes before the conversation and strongly influences every reply, setting the model’s role, tone, priorities, and boundaries without changing the model’s underlying knowledge. With this approach, the same model can sound like a professional consultant, a friendly study buddy, a pirate tutor, a strict coding mentor, or a concise expert simply by changing the instructions it receives.

It then shows how response style is affected by sampling controls, especially temperature, along with related parameters such as top_k, top_p, and repeat_penalty. Lower temperature makes outputs more predictable and conservative, while higher values increase creativity and randomness, sometimes at the cost of coherence. The chapter frames top_k as a fixed shortlist, top_p as a flexible probability-based shortlist, and repeat_penalty as a way to reduce repetitive phrasing, while noting that temperature alone is often enough for many practical applications.

Finally, the chapter brings these ideas together with Ollama Modelfiles and a Streamlit interface. A Modelfile packages a base model, system prompt, and parameter settings into a reusable custom model that can be created, listed, inspected, run, and removed locally. The Streamlit chatbot example adds controls for choosing a model, adjusting temperature, and selecting or writing personality presets, allowing users to change the chatbot’s behavior from the sidebar while the app rebuilds the request with the chosen system prompt and settings for each new message.