System prompt: the standing instructions for a chat

A system prompt is a block of instructions placed before the conversation that tells the model how to behave for the whole chat.

What a system prompt is

Chat models are trained on conversations with roles: system, user and assistant. The system message is the one meant for standing instructions such as who the model is playing, the writing style, the rules of the game, and what to avoid. Models are trained to give it extra weight compared with an ordinary user message, although how much varies a lot between models.

How it is sent

On an OpenAI-style /v1/chat/completions request it is a message with "role": "system", usually first:

{"messages": [
  {"role": "system", "content": "You are Mara, a smuggler captain. Write in third person, past tense. Never speak for the user."},
  {"role": "user", "content": "I board the ship."}
]}

On the Anthropic-style /v1/messages endpoint, the system prompt is a top-level system field instead of a message. Server-side, the model's chat template turns either form into the special tokens that model was trained on. In text completion setups, the frontend's instruct template does that formatting instead.

System prompts in SillyTavern

SillyTavern's Chat Completion presets assemble the system portion from several pieces: the Main Prompt, the character card fields, persona, lorebook entries, and optional post-history instructions placed after the chat. Cards can also carry their own system prompt that overrides the main one if you allow it. The prompt manager shows the order and lets you turn pieces on or off.

What makes a good one

• Specific over grand. "Replies of two to four paragraphs, focus on Mara's actions and dialogue" works better than "You are a world-class novelist".
• Positive instructions. Saying what to do is more reliable than a long list of bans; a ban often plants the word.
• Short. Every token is resent each turn. Long rule lists also dilute each other.
• Consistent with the rest. If the system prompt says third person but the first message is in first person, the model tends to follow the example.

A common mistake is putting critical rules only in the system prompt at the top of a 100K token chat. Instructions far from the end lose influence; an author's note or post-history instruction near the end is stronger for rules that drift.

System prompts on Wild West API

Wild West API models are uncensored, so you do not need refusal-bypass text in the system prompt. Use it for the story, the voice and the format instead. Both the OpenAI-style system message and the Anthropic-style system field are accepted on their respective endpoints.

FAQ

Related