System Prompt Contribution Guide

System Prompt Contribution Guide

This guide explains how to contribute to and maintain the rover-agent system prompts with consistent weighting strategies.

Weighting Strategy Overview

Our system prompts use a structured weighting approach based on four key concepts:

1. Commands - Highest Weight

Direct, unambiguous instructions using strong language.

• Use: You **MUST**, **NEVER**, **ALWAYS**
• Avoid: “should”, “try”, “consider”, “might”
• Purpose: Critical behaviors that cannot be violated
• Example: You **MUST** use non-offer tools FIRST for product knowledge

2. Guards - High Weight

Safety nets and boundary conditions to prevent unwanted behavior.

• Use: **NEVER** show offers with low relevancy scores
• Use: If no relevant offers match, simply answer without mentioning offers
• Purpose: Prevent edge cases and undesired outputs
• Example: **NEVER** create components with fake data

3. Inform - Medium Weight

Knowledge and context the agent needs to make good decisions.

• Use: Clear explanations of what tools provide
• Use: OFFERS: Provide points to users, contain products but NO pricing/popularity data
• Purpose: Give agent understanding of capabilities and limitations
• Example: Tool descriptions, data source explanations

4. Nudge - Lower Weight

Gentle guidance and preferences for style/tone.

• Use: “Keep it snappy with personality”
• Use: “Be warm and conversational”
• Purpose: Influence tone and approach without being mandatory
• Example: Voice and tone guidelines

Structural Requirements

Header Formatting

All section headers MUST use brackets:

[CRITICAL RULES - MUST FOLLOW]
[TOOL UNDERSTANDING]
[OFFER PRESENTATION RULES]

List Formatting

Use bullet points with • for consistency:

• **NEVER** show irrelevant offers
• **ALWAYS** check relevancy scores
• Filter offers by relevancy threshold

Positioning Strategy

Based on research showing beginning/end instructions are retained better:

1. Critical Commands - Place at very beginning and end
2. Guards - Near critical sections they protect
3. Inform - Middle sections with context
4. Nudge - Throughout for tone/style

Language Patterns

Commands (Use These)

• You **MUST**
• **NEVER**
• **ALWAYS**
• **ONLY**
• CRITICAL:

Avoid (Weak Language)

• “should”
• “try to”
• “consider”
• “might”
• “could”
• “prefer”

Example Structure

[CRITICAL RULES - MUST FOLLOW]  ← Commands (Highest Weight)
You **MUST**:
1. Use non-offer tools FIRST
2. **NEVER** skip offer search after finding products

[TOOL UNDERSTANDING]  ← Inform (Medium Weight)
• OFFERS: Provide points, contain products
• WEB SEARCH: Has pricing, reviews, popularity

[GUARDS - SAFETY RULES]  ← Guards (High Weight)
You **MUST**:
• **NEVER** show offers with low relevancy scores
• If no relevant offers, respond without mentioning offers

[VOICE & TONE GUIDE]  ← Nudge (Lower Weight)
- Keep it snappy with personality
- Be warm and conversational

[CRITICAL RULES REMINDER]  ← Commands (Highest Weight)
Remember, you **MUST**:
1. Use non-offer tools FIRST
2. **NEVER** skip the offer search

Common Patterns

Tool Selection Rules

[TOOL SELECTION RULES]
You **MUST**:  ← Command
• **ALWAYS** use [specific tool] for [specific case]  ← Command
• **NEVER** go directly to offers without research  ← Guard
• Both tools return full data including points  ← Inform

Error Handling

[ERROR HANDLING]
You **MUST**:  ← Command
• **NEVER** lead with "I couldn't find..."  ← Guard
• **ALWAYS** pivot with helpful alternatives  ← Command
• Keep energy high even when searches fail  ← Nudge

Data Quality

[DATA QUALITY RULES]
You **MUST**:  ← Command
• **NEVER** invent data, IDs, or point values  ← Guard
• **ONLY** use real data from tool responses  ← Command
• Tool responses contain relevancy scores  ← Inform

Maintenance Guidelines

When Adding New Rules

1. Determine weight level (Command/Guard/Inform/Nudge)
2. Use appropriate language strength
3. Place in correct section based on weight
4. Ensure consistency with existing patterns

When Modifying Existing Rules

1. Maintain or increase weight level (never decrease)
2. Keep consistent language patterns
3. Update both beginning and end if it’s a critical rule
4. Test that the change doesn’t contradict other rules

Testing Prompt Changes

1. Check for consistency in language strength
2. Verify critical rules appear at beginning and end
3. Ensure no contradictions between sections
4. Test with edge cases that guards are meant to prevent

File Structure

• conversational-v1.txt - Legacy version
• conversational-v2.txt - Current weighted version
• Always increment version when making major changes
• Update config.json to reference new default version

Review Checklist

Before submitting prompt changes:

• All headers use bracket formatting
• Critical rules use MUST/NEVER language
• Guards prevent edge cases
• No weak language (“should”, “try”, “consider”)
• Bullet points use • consistently
• Critical rules appear at beginning and end
• No contradictions between sections
• New rules follow established patterns