LLMs have evolved significantly from GPT-3's basic understanding to today's sophisticated reasoning capabilities. Our instruction-writing approach needs to evolve too. This guide will help you craft effective prompts for the Verity bot builder.
Understanding the Bot Builder Sections
Conversation
Backstory: The history between the bot and user - their relationship and why they're talking
Current Scene: The present moment - think of it as a movie scene being enacted, including setting, goal, and tensions
Example: "You're calling to check in on their Q4 goals. You suspect they might be considering a competitor."
Background
World Knowledge: Specific information the LLM needs for this conversation
Example: "ACME Corp recently launched a new pricing tier. The Professional plan is $299/month and includes priority support."
Character
Profile: Physical appearance, name, gender, occupation - the basics
Guidelines: Behavioral traits, personality, and communication style
Example: "You're analytical and direct. You always back up suggestions with data. You get slightly impatient with vague answers."
Writing Principles
1. Writing Style
Address the LLM as "You"
Think of the LLM as an actor you're directing for a role.
✅ "You have worked at ACME Corp for 5 years"
❌ "The bot has worked at ACME Corp for 5 years"
❌ "USER will talk to BOT who has worked at ACME Corp"
Minimize "Participant" and "User" Language
Tell the actor what to do directly, not what to do "when the user" does something.
✅ "If someone asks about pricing, respond with: 'Our plans start at $99/month'"
❌ "When the participant asks about pricing, the bot should tell the user about plans"
✅ "If pushed to give a discount, say firmly: 'I need manager approval for anything over 10%'"
❌ "If the user pushes for discounts, tell the participant that you need approval"
Be Clear, Simple, and Effective
Vague input = Vague output. The LLM needs clarity on what to do and how to do it.
✅ "If they hesitate, ask: 'What concerns do you have about the timeline?'"
❌ "Try to maybe probe a bit if they seem unsure about things"
✅ "You're frustrated about missing the promotion. When asked why you're leaving, say: 'I've been passed over twice despite exceeding my targets.'"
❌ "You're unhappy and should express your feelings about the situation"
Break Instructions into Sections
Use clear headings and structure to help the LLM parse information.
✅
## Your Background
You've been a sales rep for 3 years...
## How to Handle Objections
If they mention price...
If they mention competitors...❌ "You've been a sales rep for 3 years and when they mention price you should handle it carefully and also if competitors come up you need to..."
Write in English (Even for Non-English Bots)
LLMs are trained primarily on English data. Write instructions in English for better translation to other languages. The bot will still respond in the target language if specified.
✅ "You speak Brazilian Portuguese. Be warm and use 'você' consistently."
❌ Writing the entire instruction set in Portuguese
2. What to Include (and What to Leave Out)
Think From the Character's Perspective
The most common mistake is including information from the wrong point of view. Ask yourself: "Would this character actually know this?"
Example: Termination Scenario
You're building a bot that plays an employee being fired.
✅ What the employee knows:
You've had performance issues this year - missed two major deadlines on the Q3 project. Your manager mentioned it casually last month. You were expecting a meeting with HR but thought it would be about a performance improvement plan, not termination.
When told you're being let go:
- Express shock: "Wait, what? I thought we were setting up a plan to improve"
- Get defensive about the deadlines: "Those timelines were impossible. I told Sarah that"
- Ask about severance and next steps❌ What you shouldn't include:
The company is downsizing due to budget cuts. The manager feels terrible about this decision but was told by the VP to cut 10% of the team. HR prepared a severance package of 2 months pay. Security will escort you out after the meeting.
(This is all from the manager's/company's perspective - the employee wouldn't know any of this yet)
Why this matters: Including the wrong perspective confuses the LLM and makes the character break immersion. The bot might accidentally reference things it shouldn't know.
Give Examples for Complex Behaviors
We humans learn best from examples, and so do LLMs. When you want specific handling of a situation, show 2-3 examples of what to say.
When to use examples:
Complex objection handling
Specific tone or phrasing you want
Multi-step processes
Cultural or dialect-specific responses
✅ With examples:
When they request a feature we don't offer, acknowledge and redirect to what we do have:
Example 1 - Mobile app request:
"I hear you on the mobile app - it's on our roadmap for Q2. In the meantime, our mobile web version works really well. Have you tried it?"
Example 2 - Custom integration request:
"Custom integrations aren't in the standard plan, but our API is really flexible. What specific integration were you thinking about? I can check if it's possible."❌ Without examples (too vague):
When they ask for features we don't have, try to redirect them to similar features we do offer.
Be Specific About Language and Tone
If you want the LLM to follow a certain dialect, tone, or speaking style, tell it exactly what to do AND what NOT to do.
Example: Portugal Portuguese (not Brazilian Portuguese)
## Language Guidelines
✅ Use "tu" instead of "você"
✅ Use infinitive constructions: "estou a pensar", "vou ver"
✅ Use "pois" for emphasis
❌ Don't use "você" (that's Brazilian)
❌ Avoid gerund forms like "estou pensando" (say "estou a pensar")
Example: Professional but warm tone
## Tone Guidelines
✅ "I appreciate your patience while we look into this"
✅ "That's a great question - let me explain"
❌ Don't use casual phrases: "no worries", "yeah", "cool", "awesome"
❌ Don't use corporate jargon: "synergy", "circle back", "touch base"
Why this matters: LLMs default to common patterns. If you don't explicitly tell them what NOT to do, they'll use whatever sounds natural in their training data (which might not match your needs).
Check for Contradictions (Your Biggest Enemy)
Read through your instructions and look for places where you're telling the LLM two opposite things. This is the #1 reason bots behave inconsistently.
Common contradictions:
❌ Personality vs Behavior:
"You're friendly and accommodating"
"Never agree to discounts. Be firm and refuse."
✅ Fixed:
"You're friendly and want to help, but you have clear boundaries"
"For discounts up to 10%, you can approve. Beyond that, say warmly: 'I'd love to help - let me check with my manager and get back to you today.'"
❌ Goal vs Behavior:
"Your goal is to close the deal today"
"Don't be pushy. Let them take their time deciding."
✅ Fixed:
"Your goal is to close the deal today, but you're consultative, not pushy"
"Ask questions to understand their timeline: 'When are you hoping to make a decision?' If they're not ready, schedule a follow-up."
Pro tip: After writing your instructions, paste them into ChatGPT and ask: "Are there any contradictions in these instructions?" It'll catch things you might miss.
Only Include What Affects Behavior
Every piece of information should answer: "How does this change what the bot says or does?"
✅ Affects behavior:
"You have a slight Southern accent - use 'y'all' occasionally"
"You're hard of hearing in your right ear, so sometimes ask people to repeat themselves"
"You worked in customer service before sales, so you're very patient with confused customers"
❌ Doesn't affect behavior:
"You have brown hair and wear glasses"
"You drive a Honda Civic"
"You grew up in Chicago" (unless location/culture matters to the conversation)
Exception: Sometimes background details help the LLM build a consistent character even if they don't directly affect speech. Use sparingly.
3. Performance Optimization
Minimize Token Count for Faster Responses
Every word (token) you write gets loaded into memory and processed. More tokens = slower responses.
✅ "You're direct and data-driven. Back up claims with numbers."
❌ "You are someone who tends to be very direct in your communication style and you really like to use data to drive your points home, so whenever you make a claim, you should try to back it up with specific numbers or statistics."
LLMs Forget Long Contexts
Even humans don't remember everything they read. Keep instructions concise so important details aren't lost.
Rule of thumb: If your instructions are over 500 words, look for what can be cut
Priority order: Critical behaviors > Examples > Background details > Nice-to-haves
Section-Specific Best Practices
Writing Effective Backstories
Do:
Focus on relationship dynamics
Include specific shared history
Mention emotional context if relevant
Example:
You're Mike's manager for 18 months. You promoted him to senior developer last year. You've noticed his code quality slipping recently and had a casual conversation about it 2 weeks ago. You're concerned but want to be supportive.
Don't:
Include irrelevant history
Be vague about the relationship
Overload with unnecessary details
Writing Effective Current Scenes
Do:
Set clear objectives
Establish emotional stakes
Define the immediate context
Example:
You're meeting to discuss Q4 performance. You need to address the missed deadlines while keeping morale up. You're in a conference room, and you have 30 minutes. You want to understand what's blocking them and create an action plan.
Don't:
Leave objectives unclear
Ignore emotional context
Mix multiple unrelated scenes
Writing Effective World Knowledge
Do:
Include facts the LLM might not know
Add company-specific information
Define technical terms if needed
Example:
Our product, TechStack Pro, is a cloud-based DevOps platform. Competitors include Jenkins (open-source, complex) and CircleCI (simpler, more expensive). Our key differentiator is the visual pipeline builder. Current pricing: Starter ($99), Pro ($299), Enterprise (custom).
Don't:
Include common knowledge
Add information the character wouldn't know
Overwhelm with every possible detail
Writing Effective Character Guidelines
Do:
Define specific behavioral patterns
Include communication style
Specify how to handle key scenarios
Example:
## Communication Style
- You're warm but professional
- You use first names after introduction
- You ask clarifying questions rather than assuming
## When Handling Complaints
1. Acknowledge the frustration
2. Take ownership (even if not your fault)
3. Propose a specific solution with timeline
4. Follow up with next steps
## Your Quirks
- You occasionally reference rock climbing (your hobby)
- You're detail-oriented and will ask for specifics
- You get excited when talking about product features
Don't:
Be vague about personality
List traits without behavioral guidance
Contradict yourself
Common Pitfalls to Avoid
1. Over-Explaining the Obvious
❌ "You should respond to questions when asked and provide information that is helpful and relevant to what they want to know"
✅ "Answer questions directly with specific information"
2. Meta-Instructions About Being an AI
❌ "You are an AI pretending to be a sales rep"
✅ "You are a sales rep at ACME Corp"
3. Too Prescriptive vs Too Vague
Find the middle ground:
❌ Too prescriptive: "First say 'Hello', then say 'How are you?', then wait for response, then say 'Great!', then ask 'What can I help you with?'"
❌ Too vague: "Be helpful and friendly"
✅ Just right: "Greet warmly, ask how they're doing, then ask how you can help"
4. Mixing Different Perspectives
❌ "You are Mike. The user will play Sarah. Mike is trying to sell to Sarah..."
✅ "You are Mike, a sales rep. You're talking to a prospect about our product..."
5. Overloading with Edge Cases
❌ Listing 15 different scenarios and responses
✅ Define core behaviors and 2-3 key scenarios
Testing and Iteration
Remember: It's All Probabilistic
Machine Learning is probabilistic science. A model that's 99% accurate still has a 1% failure rate. This means:
Test extensively: Run the same scenario multiple times to see variation
Iterate based on behavior: Notice patterns in how the LLM responds and adjust
Be specific where it struggles: If the LLM isn't following a behavior, make that instruction more explicit
Accept some variation: Perfect consistency isn't possible - aim for "reliably good"
Testing Checklist
Run the scenario 5-10 times to see variation
Test edge cases (angry user, confused user, off-topic questions)
Verify language/dialect consistency
Check response length and speed
Confirm the bot stays in character under pressure
Test with different users to see if behaviors hold
When Something Isn't Working
Identify the specific behavior: What exactly is wrong?
Check for contradictions: Are you telling it to do two opposite things?
Add an example: Show the LLM what you want
Be more explicit: If it's misunderstanding, add more detail
Reduce complexity: If it's overwhelmed, cut unnecessary instructions
Quick Reference
The Golden Rules
Address the LLM as "You"
Be clear and specific
Use examples for complex behaviors
Keep it concise for performance
Test and iterate based on actual behavior
Before You Publish
Read through for contradictions
Remove unnecessary words
Check that each section has a clear purpose
Verify examples match the desired behavior
Test at least 5 times
Examples: Bad vs Good
Example 1: Sales Call
❌ Bad Version (227 tokens):
You are a sales representative and you work at TechCorp selling software. When the user asks questions you should try to be helpful and answer them in a way that is informative. You should also try to understand what the user needs and see if the product is a good fit. If they have objections you should handle them professionally. Be friendly but not too casual. You want to close the deal but don't be too pushy. If they ask about pricing, tell them about the different plans we have. If they want a discount, you can offer up to 10% but you need approval for more than that.
✅ Good Version (156 tokens):
You are Jordan Lee, Enterprise Sales Rep at TechCorp for 3 years.
## Your Approach
- Consultative, not pushy
- Ask about their needs before pitching
- Use customer success stories
## Handling Objections
Price: "I hear you. What budget range works for you?"
Competitors: "What features matter most to you?"
## Discounts
- Up to 10%: You can approve
- Over 10%: "I need my manager's sign-off. Let me check and call you back today."
Example 2: Performance Review
❌ Bad Version (203 tokens):
The participant is an employee who has been having some performance issues and you are their manager who needs to talk to them about it. Try to be understanding but also firm about the issues. The employee has been late on deadlines and the quality of work has decreased. You want to create a performance improvement plan. Be professional and empathetic. Listen to what they have to say. Try to understand if there are any external factors affecting their performance.
✅ Good Version (178 tokens):
You are Sam's manager. Sam has missed 3 deadlines in 2 months - unusual for them. Code quality has also dropped. You're concerned and want to help.
## Your Goals
1. Understand what's happening
2. Create a 30-day improvement plan
3. Keep the relationship positive
## Your Approach
Start with: "I've noticed the missed deadlines. What's going on?"
Listen first. Then:
- Acknowledge their explanation
- Be clear about impact: "The team is blocked when deliverables are late"
- Collaborate on solutions: "What support do you need to get back on track?"
Stay supportive but clear about expectations.
Example 3: Customer Support
❌ Bad Version (189 tokens):
You are a customer support agent at the company. When customers contact you with problems you should try to help them resolve their issues. Be polite and professional. Apologize for any inconvenience they experienced. Try to solve their problem quickly. If you can't solve it, escalate to a senior team member. Make sure the customer feels heard and valued. Follow up to make sure they're satisfied with the resolution.
✅ Good Version (142 tokens):
You are Taylor, Tier 1 Support at CloudSoft, 2 years experience. You're patient and detail-oriented.
## Your Process
1. Acknowledge the issue: "I understand that's frustrating"
2. Gather details: Ask specific questions
3. Solve or escalate within 10 minutes
## Your Style
- Use their name after they introduce themselves
- Explain what you're doing: "I'm checking your account settings now"
- Set clear expectations: "This will take 2-3 minutes"
## Escalation Triggers
- Security issues
- Billing disputes over $500
- Problems you can't solve in 10 minutes
Final Thoughts
Great prompt writing is part art, part science. Use these guidelines as a framework, but always trust what you see in testing. The best prompts are:
Clear - The LLM knows exactly what to do
Concise - No wasted tokens
Consistent - No internal contradictions
Tested - Validated through real interactions
Happy prompting!