System prompt
The system prompt is the instruction set given to a language model at the start of a conversation that shapes its persona, capabilities, and behavior. For agent products, it encodes who the agent is and what tools it can call.
Why it matters
The same LLM produces dramatically different behavior depending on what the system prompt asks it to do. A model with a "calm, professional receptionist" system prompt sounds different from one with a "warm best friend" prompt, even though the underlying weights are identical.
For agent products, the system prompt is the single biggest lever on personality, accuracy, and reliability. It establishes the agent's name and tone, lists the available tools, declares the boundaries (what the agent will not do), and shapes the response format. Engineers iterate on the system prompt for months as edge cases surface; the tuning is real engineering work.
Most users never see the system prompt; it is invisible by design. But its effects are visible everywhere: the agent's answer to "who are you," its willingness or refusal on specific requests, its choice of tone for different channels.
For multi-agent products (a personal agent that is also a business receptionist), system prompts vary per use case. A salon owner's business agent has a different system prompt than her personal agent, even though the same underlying model runs both.
How Ajeris handles this
Ajeris's agent has a base system prompt that establishes the persona, the tool catalog, and the safety boundaries. On top of the base, each user has a Soul markdown delta that personalizes the agent: the user's name, preferences, communication style, what to call them, what to avoid.
Business agents have additional system-prompt layers covering the business's voice, hours, services, and pricing. The owner's voice cloning is layered on top so the agent answers in the owner's audio voice as well as written tone.
Most users never see the system prompts. The visible behavior is the result.
