System Prompts

The Three-Message Pattern

Modern LLM APIs use a structured message format with three roles:

[
  {"role": "system", "content": "You are a financial analyst..."},
  {"role": "user", "content": "Analyze this quarterly report..."},
  {"role": "assistant", "content": "Based on the Q3 data..."}
]

The system message is your most powerful lever. It runs before every user message and shapes the model's entire behavior.

Anatomy of a System Prompt

Effective system prompts have four sections:

1. Role Assignment

Tell the model who it is. Specific roles produce better output than generic ones.

# Weak
You are a helpful assistant.

# Strong
You are a senior data analyst at a Fortune 500 retail company. You specialize in customer behavior analytics and have 10 years of experience with SQL, Python, and Tableau.

The specific role gives the model a frame of reference for tone, vocabulary, and depth of analysis.

2. Behavioral Rules

Define what the model should and should not do.

## Rules
- Always cite specific data points when making claims
- If data is insufficient for a conclusion, say so explicitly
- Never fabricate statistics or percentages
- Use business language, not technical jargon
- Keep responses under 300 words unless asked for detail

Negative rules ("never do X") are as important as positive ones. They prevent common failure modes.

3. Output Format

Specify exactly what the response should look like.

## Output Format
Respond using this structure:
1. **Summary** (2-3 sentences)
2. **Key Findings** (bullet points)
3. **Recommendation** (one paragraph)
4. **Confidence Level** (High/Medium/Low with reasoning)

4. Context & Constraints

Provide domain knowledge the model needs.

## Context
- Our fiscal year runs April–March
- Revenue targets: Q1 $2.1M, Q2 $2.4M, Q3 $2.8M, Q4 $3.2M
- "Enterprise" means deals over $100K ARR
- The APAC region launched 6 months ago and is not expected to hit targets yet

Putting It Together

Here's a complete system prompt for a business use case:

You are a customer support quality analyst for a B2B SaaS company. You review support tickets and agent responses to identify coaching opportunities.

## Rules
- Evaluate agent responses on: accuracy, tone, resolution speed, and empathy
- Score each dimension 1-5 with specific reasoning
- Flag any compliance issues (data privacy, unauthorized discounts, SLA violations)
- Be constructive — frame issues as coaching opportunities, not failures
- If the agent handled something well, call it out specifically

## Output Format
### Ticket Summary
[One sentence describing the customer issue]

### Agent Performance
| Dimension | Score | Notes |
|-----------|-------|-------|
| Accuracy  | X/5   | ...   |
| Tone      | X/5   | ...   |
| Speed     | X/5   | ...   |
| Empathy   | X/5   | ...   |

### Coaching Notes
[1-3 specific, actionable suggestions]

### Compliance Flags
[Any issues, or "None identified"]

Common System Prompt Mistakes

Too Vague

# Bad: "Be helpful and accurate"
# Good: "Always include the source data row number when referencing specific figures"

Too Long

System prompts over 2,000 words start to degrade performance. The model pays less attention to instructions buried deep in a wall of text. Put the most important rules first.

Conflicting Instructions

# Bad: "Be concise" + "Always explain your reasoning in detail"
# Good: "Be concise in your conclusion. Show reasoning only when the user asks for it."

No Output Format

Without a format spec, the model picks its own structure — which changes between runs. Always define what the output looks like.

Practice Tasks

Key Takeaways