Skip to main content

Prompts Overview

Learn how system prompts control the chatbot's behavior, tone, and response generation, and why careful prompt management is critical for chatbot quality.

What Are System Prompts?​

Simple Explanation​

System prompts are instructions given to the AI that tell it how to behave when answering user questions.

Think of it like:

  • An employee handbook that defines company policies
  • A script that actors follow in a play
  • Rules that a referee enforces in a game

What prompts control:

  • 🎭 Tone: Friendly, formal, professional, casual
  • 📝 Style: Concise, detailed, academic, conversational
  • 🎯 Focus: What topics to prioritize or avoid
  • ⚖️ Rules: What the chatbot can and cannot do
  • 🔍 Behavior: How to handle uncertainty, errors, or edge cases

Example: How Prompts Change Behavior​

Without a system prompt, the AI might respond generically:

User: "What is HFIM?"

Response: "HFIM typically stands for various things depending on context.
Could you provide more information about what you're asking about?"

With a good system prompt, the AI responds specifically and helpfully:

User: "What is HFIM?"

Response: "HFIM stands for Hospitality and Food Industry Management, a degree
program at the University of Georgia that prepares students for careers in
hospitality, tourism, and food service management. The program combines
business fundamentals with hands-on industry experience."

The difference: The system prompt told the AI "You are a UGA HFIM expert assistant. Always interpret HFIM as Hospitality and Food Industry Management."

Why Prompts Matter​

Impact on User Experience​

Good prompts create:

  • ✅ Consistent, professional responses
  • ✅ Accurate, on-topic answers
  • ✅ Appropriate tone for the audience
  • ✅ Clear, well-structured content
  • ✅ Transparent handling of uncertainty

Bad prompts create:

  • ❌ Inconsistent responses (tone varies wildly)
  • ❌ Off-topic or irrelevant answers
  • ❌ Inappropriate tone (too casual or too formal)
  • ❌ Confusing or poorly structured content
  • ❌ Made-up information (hallucinations)

Real-World Example​

Scenario: User asks "How hard is the HFIM program?"

Without prompt guidance:

Response: "The HFIM program can be challenging, but it depends on your
background and work ethic. Some students find it difficult while others
don't. It varies from person to person."

Vague, unhelpful, doesn't provide actionable information

With prompt guidance ("Always provide specific examples and data when possible"):

Response: "The HFIM program is moderately rigorous, combining business
coursework with hands-on hospitality training. Students typically:

• Complete 120 credit hours over 4 years
• Take courses in accounting, marketing, operations management
• Maintain a 2.5 GPA to stay in the major
• Complete a 400-hour internship requirement

Students report that time management and strong study habits are key to
success. The workload is comparable to other business majors at UGA."

Specific, actionable, addresses the question thoroughly

Types of Prompts in the Admin Panel​

1. System Prompt (Main Prompt)​

What it is: The primary set of instructions given to the AI for all responses

Controls:

  • Identity ("You are a UGA HFIM assistant")
  • Core rules (don't make up information, always cite sources)
  • Response format (use bullet points, be concise)
  • Handling edge cases (what to do when uncertain)

Location: Prompts Configuration → System Prompt tab

Frequency of updates: Rarely (2-4 times per year)

2. Reframing Prompt (Advanced)​

What it is: Instructions for reformulating unclear user questions before searching

Purpose: Improve search quality by clarifying vague questions

Example:

  • User asks: "whos prof for 3000"
  • Reframing prompt helps AI interpret as: "Who teaches HFIM 3000?"

Location: Prompts Configuration → Reframing Prompt tab (if available)

Frequency of updates: Rarely (only when search quality issues arise)

3. Custom Prompts (If Available)​

What they are: Specialized prompts for specific scenarios

Examples:

  • Welcome message prompt
  • Error handling prompt
  • Follow-up question prompt

Location: Depends on admin panel configuration

Frequency of updates: As needed for specific use cases

How Prompts Work​

The Response Generation Process​

Step-by-Step:

  1. User Asks Question

    • "What are the prerequisites for HFIM 3000?"

  2. System Loads Prompts

    • System prompt: "You are a UGA HFIM assistant..."
    • Reframing prompt (if needed): "Clarify vague questions..."

  3. Search for Information (if not cached)

    • Search Pinecone for relevant documents
    • Retrieve top 5-10 most relevant chunks

  4. Combine Prompt + Context + Question

    • System prompt (behavior rules)
    • Retrieved documents (knowledge)
    • User question (what they want to know)

  5. AI Generates Response

    • Follows rules in system prompt
    • Uses information from documents
    • Answers user's specific question

  6. Response Returned to User

    • Formatted according to prompt guidelines
    • Cites sources as instructed in prompt

What You Can (and Can't) Control with Prompts​

What Prompts CAN Control​

✅ Tone and Style

  • Formal vs. casual language
  • Professional vs. friendly personality
  • Concise vs. detailed responses

✅ Response Structure

  • Use of bullet points
  • Paragraph length
  • Section headings

✅ Rules and Constraints

  • Never make up information
  • Always cite sources
  • Admit uncertainty when appropriate

✅ Topic Focus

  • Prioritize HFIM-related content
  • Emphasize certain aspects (career paths, courses, etc.)

✅ Error Handling

  • What to say when uncertain
  • How to handle off-topic questions
  • Suggestions for rephrasing

What Prompts CANNOT Control​

❌ Factual Knowledge

  • Prompts don't add information to the AI's knowledge base
  • AI only knows what's in the ingested documents

❌ Real-Time Information

  • Prompts can't make the AI aware of current dates/events
  • Use cache entries for date-specific information

❌ User-Specific Context

  • Prompts don't give AI access to user's personal info
  • Each conversation is anonymous

❌ Search Algorithm

  • Prompts don't change which documents are retrieved
  • Search quality is separate from prompt quality

Prompt Best Practices Overview​

Core Principles​

1. Be Specific and Clear

  • ✅ "Always cite sources using this format: [Document Name, Page X]"
  • ❌ "Cite sources"

2. Set Explicit Rules

  • ✅ "If you're uncertain about an answer, say 'I don't have enough information to answer this question accurately.'"
  • ❌ "Try to answer questions"

3. Provide Examples

  • ✅ "When asked about courses, respond like this: [example]"
  • ❌ "Answer questions about courses"

4. Maintain Consistency

  • ✅ "Always use 'HFIM' (not 'Hospitality program' or 'HFIM program')"
  • ❌ Use varied terminology

5. Test Before Deploying

  • ✅ Preview changes with sample questions
  • ❌ Save changes without testing

Common Prompt Mistakes​

Mistake 1: Vague Instructions​

Bad: "Be helpful"

Good: "Provide specific, actionable information. Use examples and numbers when possible. Break complex information into bullet points."

Mistake 2: Contradictory Rules​

Bad:

- Always be concise (under 50 words)
- Always provide detailed explanations with examples

These conflict - can't be both concise and detailed

Good:

- Provide concise summaries (50-100 words) with option to expand
- Use bullet points for easy scanning
- Include examples only when they add clarity

Mistake 3: Over-Constraining​

Bad:

- Responses must be exactly 75 words
- Always use exactly 3 bullet points
- Never use the word "very" or "really"
- Begin every response with "Great question!"

Too many rigid rules make responses robotic and awkward

Good:

- Keep responses between 50-150 words
- Use bullet points for lists (3-5 items typically)
- Use strong adjectives instead of intensifiers ("excellent" vs. "very good")
- Acknowledge questions naturally

Mistake 4: Forgetting to Test​

Bad: Save prompt changes and assume they'll work

Good:

  1. Edit prompt
  2. Preview with 5-10 sample questions
  3. Verify responses are appropriate
  4. Save only after testing

Prompt Change Impact​

Hot Reload System​

What it means: Changes to prompts take effect immediately (within seconds)

Impact:

  • ⚠️ No "undo" button - changes go live instantly
  • ⚠️ All users see the new behavior immediately
  • ⚠️ Bad changes affect everyone until you fix them

Why this matters:

  • Test carefully before saving
  • Monitor conversations after changes
  • Have a rollback plan (use Change History)

Measuring Prompt Changes​

After changing a prompt, monitor:

  1. User Feedback (next 24 hours)

    • Check for increase in negative feedback
    • Read feedback comments for issues

  2. Response Quality (next 3-7 days)

    • Spot-check 10-20 conversations
    • Verify responses match your intent

  3. Common Issues (next 1-2 weeks)

    • Watch for patterns (formatting problems, tone issues)
    • Adjust prompt if needed

If issues arise: Use Change History to revert to previous version

Getting Started with Prompts​

For New Admins​

Your first month:

  1. Week 1: Read current prompts (don't change anything)
  2. Week 2: Observe user conversations, note any tone/style issues
  3. Week 3: Propose minor prompt adjustments to team
  4. Week 4: Make small, tested changes if approved

Don't rush: Prompt changes are powerful and affect all users. Learn before editing.

For Experienced Admins​

Recommended schedule:

  • Monthly: Review prompt performance based on feedback
  • Quarterly: Audit entire prompt for outdated info or conflicting rules
  • As needed: Update prompts when program requirements change

Keep a changelog: Document why you made each change for future reference

Next Steps​

Now that you understand prompts:

  1. Learn about system prompts - Understand the main prompt structure
  2. Learn to edit prompts safely - Step-by-step editing guide
  3. Understand change history - Track and revert changes
  4. Learn testing practices - Validate changes before saving
  5. Review hot reload - Understand immediate deployment
  6. Read best practices - Proven strategies for prompt management

Remember: Prompts are the chatbot's "brain" - they control everything. With great power comes great responsibility. Test carefully and monitor closely!

Last updated on