teaching_material — System Prompt

So far, we've used simple system prompts like "You are a helpful assistant" or "You are a biology tutor." But in real applications — especially agents that use tools and need to follow complex workflows — system prompts become much more detailed and structured. The Biomni agent (a biology research assistant built at Stanford) is a good example of a production-grade system prompt. By studying it, you'll see how professional LLM applications structure their instructions to ensure consistent, reliable behavior. This section is optional, but highly recommended if you plan to build your own agents in later lectures.

A well-designed system prompt typically includes: • Role Definition — Who is the agent? What is its purpose? • Example: "You are Biomni, a biology research assistant specializing in molecular biology and drug discovery." • Behavioral Guidelines — How should the agent respond? • Example: "Always cite sources," "Explain technical concepts in simple terms," "Admit when you don't know something." • Available Tools (for agents) — What actions can the agent take? • Example: "You have access to 50+ tools including PubMed search, protein structure lookup, and pathway analysis." • Output Format — How should responses be structured? • Example: "Format tool calls as JSON," "Use markdown for formatting," "Limit responses to 200 words." • Constraints and Rules — What should the agent NOT do? • Example: "Never make up data," "Don't perform medical diagnosis," "Stay within the biology domain."

The Biomni agent uses a ~2-page system prompt that exemplifies all the components above. You can find it in the Biomni paper (search for "Biomni prompt" — it appears in the supplementary materials). What to look for when reading it: • Notice how the role is defined in the first paragraph • Count how many tools are listed and how they're described • Identify the output format requirements (JSON structure, citation format) • Look for constraints — statements with "always," "never," or "must"

Here's how the Biomni prompt compares to what we've used so far: | Our prompts | Biomni prompt | |-------------|---------------| | 1 sentence | ~2 pages | | "You are a helpful biology tutor" | Detailed role with specialty areas | | No tools | 50+ tools with descriptions | | No format requirements | Strict JSON output format | | No explicit constraints | 10+ behavioral rules | When do you need a complex prompt like Biomni's? • When building agents that use tools (Lecture 4–5) • When consistency across many requests is critical • When working in regulated domains (medical, legal) • When you need to constrain model behavior tightly When is a simple prompt enough? • One-off queries or simple Q&A • Exploratory data analysis • Non-critical applications where variability is acceptable