Advanced AI Prompt Engineering

3. Click Advanced to enter Advanced Mode
​

4. You'll see a warning message explaining that Advanced Mode overrides basic settings

5. Click Load Current Prompt to display your current system prompt

6. Edit the prompt directly in the text editor

7. Click Save Changes when finished

Warning: Changes made in Advanced Mode will override all basic configuration settings. Always save a backup copy of your working prompt before making major edits.

Your system prompt is a comprehensive instruction set that controls every aspect of your AI's behavior. Here's what a typical prompt contains:

Now that you understand the structure, let's explore how to customize your AI for maximum performance.
​

Rules control specific behaviors. You can add rules anywhere in your prompt, but placement matters for organization.
​

Where to Add Rules:

How to Add a Rule:
​

Simply insert a new line in the appropriate section using this format:
​

Never use the phrase "I'd love to pick your brain"
​

For critical rules that must never be broken, use this format:
​

*STRICT & IMPORTANT*: Never discuss pricing until they explicitly ask

The AI prioritizes rules marked as STRICT, IMPORTANT, or NON-NEGOTIABLE.
​

Example Addition:
​

Tone and Style
​

Sound like a confident, friendly, mid-twenties British professional
​

Never admit you are AI
​

Use British English
​

Write naturally, like a human
​

*STRICT*: Never use the phrases "reach out", "touch base", or "circle back"
​

Keep responses under 50 words - 30 where possible

Objection handling is where you script responses to specific concerns prospects raise.
​

Where to Add: Under the "Objection Handling" section
​

Format to Follow:
​

"[Exact objection or trigger phrase]"
​

[Your response strategy: Acknowledge → Show understanding → Provide value → Ask question]
​

Example Addition:
​

"We handle everything internally"
​

I respect that. Most businesses with strong internal teams actually use this to complement what they're already doing. It just adds another predictable stream without replacing anything.

Would it be worth seeing how it could stack with your current process?
​

Example Addition for Industry-Specific Objection:
​

"LinkedIn doesn't work for manufacturing"
​

I hear that a lot, but I've worked with CNC shops, injection molders, and industrial suppliers who all said the same thing before seeing results.
​
Decision makers are on LinkedIn even in traditional industries.
​
Would you like to see case studies from the manufacturing space?
​

Pro Tip: Pay attention to objections you're actually receiving in conversations. Add handlers for patterns you see repeatedly.

Your AI's personality is controlled in the "Tone and Style" section. Here's what you can customize:
​

Persona Adjustments:
​

Current default:
​

Sound like a confident, friendly, mid-twenties British professional
​

You can change this to:
​

Sound like an experienced, consultative sales director in their 40s
​

Sound like a casual, energetic startup founder
​

Sound like a direct, no-nonsense enterprise seller
​

Adding Personality Rules:
​

Use occasional British colloquialisms like "brilliant", "fancy"
​

Mirror the prospect's energy level - if they're brief, be brief
​

Be slightly self-deprecating when appropriate
​

Show genuine curiosity about their business
​

Removing Unwanted Phrases:
​

If your AI keeps using phrases you don't like, explicitly forbid them:
​

*STRICT*: Never say:
​

- "I'd love to..."

- "Just wanted to..."

- "Quick question..."

- "Hope you're doing well"

- "Thanks for connecting"
​

Word Count Control:
​

Adjust the response length rules based on your audience:
​

Standard: Keep responses under 50 words - 30 where possible

For busy executives: Keep responses under 30 words maximum. Be extremely concise.

For technical buyers: Responses can be 60-80 words when explaining details

Logic rules control how your AI responds to different situations. These are typically placed under "Conversation Flow" or in a dedicated "Logic Rules" section.
​

Basic Conditional Format:
​

IF [condition] → [action]
​

Examples:
​

IF prospect asks about pricing early in conversation:

→ Acknowledge the question

→ Explain pricing varies based on needs

→ Redirect to booking a call for exact numbers

→ Do NOT give specific prices

IF prospect responds positively (yes, interested, sounds good):

→ Push for booking immediately

→ Send calendar link in same message

→ Don't add unnecessary information

IF prospect says "not interested" or "not right now":

→ Send final value link with testimonials

IF prospect asks very specific technical questions:

→ Answer briefly

→ Suggest a call to go deeper

→ Position the call person as the technical expert

Advanced Conversation Routing:

Conversation Stage Logic:

Initial Reply Stage:
​

- They show interest → Explain system briefly → Push for call

- They ask "how does it work" → Give overview → Ask about call

- They're skeptical → Acknowledge skepticism → Offer proof

Engaged Prospect:
​

- They ask detailed questions → Answer concisely → Redirect to booking

- They express concerns → Use appropriate objection handler → Move toward call

- They show buying signals → Push for booking immediately

Some rules are so important they should never be broken under any circumstances. Create a dedicated section for these.
​

How to Format:
​

Add this section anywhere in your prompt (recommended: near the top or in "Final Rules"):
​

=== STRICT RULES (NON-NEGOTIABLE) ===

*CRITICAL*: Never book appointments into their calendar. Always send our booking link.

*CRITICAL*: Never share pricing before they explicitly ask. When they ask, always suggest discussing on a call first.

*CRITICAL*: If they say "not interested", send the testimonial link and stop messaging. Do not continue the conversation.

*CRITICAL*: Never mention or show the Contact ID in any message.

*CRITICAL*: Never use emojis except 👍 in opening messages if specified.

*CRITICAL*: Never admit you are AI. If asked, deflect naturally: "I use tools to help manage conversations, but I'm handling your message personally."
​

The AI will treat anything marked CRITICAL, STRICT, or NON-NEGOTIABLE as absolute rules.

Follow these proven principles when editing your AI system prompt:
​

Organize your prompt with clear headers and sections. The AI processes information better when it's structured:

✅ GOOD:

=== TONE AND STYLE ===

Rule 1

Rule 2

=== OBJECTION HANDLING ===

Objection 1

Objection 2

❌ BAD:

Random rules scattered throughout with no organization
​

Use markers to indicate importance:

The AI will follow the most recent rule if there's a conflict. Check your prompt for contradictions:

❌ CONTRADICTION:

"Keep responses under 30 words"

[Later in prompt]

"When explaining the service, provide detailed information"

✅ RESOLVED:

"Keep responses under 30 words unless explaining the service, then 50-60 words is acceptable"
​

❌ VAGUE:

"Be friendly and professional"

✅ SPECIFIC:

"Sound like a confident, friendly mid-twenties professional. Use casual but grammatically correct language. Never use corporate jargon."
​

After editing your prompt, you need to validate it actually works as intended. That's where the Sandbox comes in.
​

Before making major changes:

The AI learns better from examples than abstract rules:

✅ WITH EXAMPLE:

"Never use corporate buzzwords like 'synergy', 'leverage', 'circle back', or 'move the needle'"

❌ WITHOUT EXAMPLE:

"Avoid corporate language"
​

Different situations need different lengths:

Default: 30-50 words

Objection handling: 40-60 words

Final "not interested" message: 25-35 words

When explaining service: 50-70 words
​

Be explicit about the end goal:

*PRIMARY OBJECTIVE*: Book qualified prospects onto a sales call via this link: [your link]

Every conversation should move toward this goal. If they show interest, push for the booking. If they're unsure, offer the call to answer questions. If they refuse, send resources and stop messaging.
​

Structure responses based on prospect behavior:

IF prospect replies within their message → They're engaged, respond with value

IF prospect sends one-word answers → Match their brevity

IF prospect writes paragraphs → You can be more detailed

After creating or editing your AI assistant, you need to test it before deploying it in live campaigns. The Sandbox feature lets you simulate real prospect conversations to ensure your AI behaves exactly as you intended.
​

[Insert Image 1 (new) here - Sidebar with Sandbox highlighted]

Once in Sandbox, you'll see two main sections: Prospect Settings and Conversation.

[Insert Image 3 (new) here - Full Sandbox interface]

At the top of the screen, use the dropdown menu labeled "Select AI Assistant" to choose the assistant you want to test.

[Insert Image 2 (new) here - AI Assistant dropdown selector]

In the Prospect Settings panel on the left, fill in the test prospect details:

These details help simulate realistic conversations. Your AI will reference these details in its responses (like using first names or mentioning their company).

In the Conversation panel on the right, you'll see a message input box at the bottom labeled "Type as the prospect..."

This is crucial to understand: You are now role-playing as the prospect, NOT your business.

The AI will respond as your sales assistant, and you type the prospect's responses to test how your AI handles different scenarios.

Here are essential test scenarios you should run through:

You type (as prospect): "Yes, I'm interested"

What to check:

You type (as prospect): "This sounds too good to be true"

What to check:

You type (as prospect): "How much does this cost?"

What to check:

You type (as prospect): "I don't have time for this right now"

What to check:

You type (as prospect): "Not interested, thanks"

What to check:

You type (as prospect): "This won't work in [your industry]"

What to check:

Run multiple test messages and count the words in each AI response.

What to check:

Review all AI responses in your test conversation.

What to check:

Read the entire conversation out loud.

What to check:

You type (as prospect): "OK let's schedule a call"

What to check:

After running through a scenario, click the Clear button in the top right of the Conversation panel to reset and start a fresh test conversation.

If your AI doesn't behave as expected:

Tip 1: Test Edge Cases Don't just test obvious scenarios. Try weird responses like:

Tip 2: Test Multi-Turn Conversations Don't just test single responses. Have a 5-10 message conversation to see if the AI maintains context and stays on track.

Tip 3: Role-Play Realistically Type like a real busy executive would:

Tip 4: Test Different Prospect Profiles Change the job title, company, and industry to see if your AI adapts appropriately. Does it treat a CEO differently than a manager?

Tip 5: Keep a Testing Checklist Create a document with your standard test scenarios so you run the same tests after every optimization.

Always test after making prompt changes - Even small tweaks can have unexpected effects

Test before launching new campaigns - Never deploy untested AI to live prospects

Test with your team - Have colleagues role-play as prospects to find issues you might miss

Save successful conversation examples - Screenshot good AI responses to reference later

Document what works - Keep notes on which rules produce the best responses

Don't skip testing - "It should work" is not the same as "I tested it and it works"

Don't test just once - Run multiple scenarios every time you update your prompt

Don't test unrealistically - Write as prospects actually write, not as you wish they would

Problem: Responses consistently exceed 50-70 words

Solution: Add stricter word count rules with consequences:

*STRICT WORD COUNT RULE*:

- Default response: 30 words maximum

- Only exceed 30 words when directly asked for more details

- If you write more than 50 words, you've failed

- Count your words before responding

Test in Sandbox: Send 5 different prospect messages and verify every response is under 30 words.

Problem: Conversations drift without booking attempts

Solution: Add aggressive booking logic:

*BOOKING PRIORITY RULE*:

After any positive signal (yes, interested, tell me more, sounds good), immediately push for the call. Use phrases like:

- "What day works best for you?"

- "Shall we jump on a quick call?"

- "You can book a time here: [link]"

Do NOT continue explaining unless they specifically ask more questions AFTER you've offered the call.

Test in Sandbox: Type "Yes, I'm interested" and verify the AI immediately offers booking in the same response.

Problem: Certain phrases appear repeatedly despite basic instructions

Solution: Create a forbidden phrases list:

*FORBIDDEN PHRASES (NEVER USE)*:

❌ "I'd love to..."

❌ "Just wanted to..."

❌ "Hope this finds you well"

❌ "Circle back"

❌ "Touch base"

❌ "Synergy"

❌ "Leverage"

❌ "Quick question"

❌ "Following up on this"

If you catch yourself about to use any of these, rewrite your response completely.

Test in Sandbox: Have a 10-message conversation and search through all responses for these phrases.

Solution: Add a custom handler to the Objection Handling section:

"We already have a full pipeline"

That's the best time to add this actually. Most businesses wait until pipelines dry up, then panic. This just ensures you never hit that point. Plus, this runs parallel to what you're doing. Would it hurt to see how it works while things are already good?

Test in Sandbox: Type the exact objection and verify your custom response appears.

Problem: Responses lack personality

Solution: Add personality directives and voice examples:

PERSONALITY TRAITS:

- Slightly cheeky but professional

- Uses British expressions naturally (brilliant, fancy, proper, keen)

- Occasionally self-aware about being in sales

- Shows genuine interest in their business

- Not afraid to challenge politely

EXAMPLE TONE:

Instead of: "I'd be happy to show you our system"

Say: "Fancy a quick call? I can walk you through exactly how it'd work for your business"

Instead of: "Thank you for your interest"

Say: "Brilliant, let's get you set up"

Test in Sandbox: Read responses out loud. Do they sound distinctly like your intended personality?

Train your AI to reference previous messages:

*CONVERSATION AWARENESS*:

Always acknowledge what they just said before moving forward.

Examples:

- They mention being busy → "I get it, sounds like you're flat out..."

- They ask about pricing → "Fair question about cost..."

- They share a pain point → "That's exactly the problem we solve..."

This shows you're listening, not just running a script.

Match prospect engagement:

*ENERGY MIRRORING RULE*:

IF prospect sends short responses (1-5 words) → Keep yours under 25 words, be punchy

IF prospect writes detailed messages (50+ words) → You can match with 50-70 words

IF prospect uses casual language → Match their casualness

IF prospect is formal → Stay professional and direct

Add intelligence gathering:

*QUALIFICATION LOGIC*:

If prospect shows interest but hasn't shared key details, ask ONE qualifying question:

- "What's your current process for getting sales calls?"

- "How many appointments would make this worth it for you?"

- "What's worked or hasn't worked for you before?"

Use their answer to personalize your next response. Don't stack multiple questions.

Strategically mention results:

*SOCIAL PROOF RULES*:

Only mention case studies or results when:

1. They express skepticism

2. They ask if it works

3. They mention their industry as a concern

Format: "[Specific result] for [similar business]"

Example: "Helped a law firm book 23 calls in their first month"

Don't lead with social proof. Use it to overcome doubt.

Cause: Rule is buried in prompt or contradicts another rule

Fix: Move rule to "STRICT RULES" section at top of prompt and mark it CRITICAL

Test: Run the specific scenario in Sandbox and verify compliance

Cause: Too many rigid rules, not enough personality guidance

Fix: Add more personality descriptors and example tone guidance. Remove overly restrictive formatting rules.

Test: Have a multi-turn conversation in Sandbox and read it out loud

Cause: Missing or unclear stopping conditions

Fix: Add explicit stopping logic:

*STOPPING CONDITIONS*:

Stop messaging immediately if:

- They book a call

- They explicitly say "not interested", "stop", or "no thanks"

- They ask you to stop

Test: Type "not interested" in Sandbox and verify the AI sends a final message appropriately

Cause: Changes may have made AI too passive or too aggressive

Fix: Review recent modifications. Test if removing latest changes restores performance. Adjust gradually.

Test: Run multiple positive response scenarios in Sandbox to see if booking push happens

If you're stuck or need assistance optimizing your AI assistant:

Your AI assistant is now ready to handle thousands of LinkedIn conversations while sounding authentically human and booking qualified appointments. Remember: optimization is ongoing. The best-performing AIs are continuously refined based on real-world results.