In Part 1 of this series, we established the strategic value of synthetic users: they solve the “Cold Start” problem by allowing you to simulate 100 customer interviews overnight.
But strategy is cheap. Execution is where most founders fail.
If you open ChatGPT orClaude right now and type: “Act like a Fortune 500 CISO and tell me if you would buy my software,” you will get a useless answer.
You will get a polite, generic hallucination. The AI will likely say, “As a CISO, I prioritize security and scalability. Your tool looks promising!”
This is the “Internet Consensus” Trap. Without specific constraints, Large Language Models (LLMs) revert to the average of their training data. They give you the “average” answer that an “average” CISO would give on the “average” day.
In B2B SaaS, you don’t sell to averages. You sell to specific people with specific problems.
To get value from synthetic research, you must move beyond prompting and start engineering. You need to build High-Fidelity Personas—digital twins that mirror the messy, irrational, and specific constraints of your actual buyers.
Garbage In, Garbage Out: Why Simple Prompts Fail
The difference between a toy and a tool is Grounding.
A generic prompt relies on the LLM’s internal weights (what it “read” on the internet). A grounded prompt forces the LLM to rely on your data.
If you are building an AI Generated ICP, you know that a real buyer isn’t just a job title. They are a collection of anxieties, budget constraints, and tech stack limitations.
- Generic Prompt: “Act like a Marketing Manager.”
- Result: A caricature who talks about “brand synergy” and “engagement.”
- Engineered Prompt: “You are a Performance Marketer at a Series B Fintech. You are under pressure to lower CAC by 20% this quarter. You use HubSpot and Salesforce, and they don’t sync properly. You hate new tools that require engineering support.”
- Result: A stressed decision-maker who rejects your tool because it lacks a native Salesforce integration.
The second result saves you 6 months of development time. The first result wastes it.
The Persona Architecture: The 4-Part System Prompt
To build a high-fidelity synthetic user, you need to structure your System Prompt like a character sheet in a role-playing game.
We use a 4-part architecture for our AI-Accelerated PMF Validation:
1. Demographics & Firmographics (The “Who”)
This is the easy part. Define the shell.
- Role: VP of Engineering.
- Company: 500-1,000 employees, Healthcare sector.
- Revenue: $50M – $100M ARR.
- Location: North America (implies GDPR/CCPA relevance).
2. Psychographics (The “Why”)
This is where the magic happens. You need to define their internal state.
- Motivation: “I am risk-averse. My predecessor got fired for a data breach.”
- Bias: “I prefer ‘boring’ established vendors (Microsoft, Oracle) over startups.”
- Personality: “Direct, skeptical, hates buzzwords, short attention span.”
3. Context Grounding (The “World”)
You must anchor them in reality. This prevents hallucinations about tools that don’t exist.
- Tech Stack: Uses AWS, Jira, and Slack.
- Pain Points: “Drowning in technical debt,” “Hiring freeze in effect.”
- Source Data: Tip: Paste 3-5 anonymized reviews of your competitors from G2 or Capterra here. Tell the AI: “This is how you speak about software.”
4. The Mission (The “What”)
Never leave the AI open-ended. Give it a job.
- Wrong: “Chat with me.”
- Right: “Critique this landing page specifically for compliance risks. If you see a claim that seems unprovable, flag it aggressively.”
The “Anti-Persona”: Engineering for Dissent
One of the biggest risks in AI research is Sycophancy—the AI wanting to please the user. We wrote extensively about this in our guide on AI User Research Bias.
To counter this, you must engineer an “Anti-Persona.”
If you are selling a premium tool, do not just build “Ideal Customer Ian.” Build “Budget-Hawk Bob.”
- Instruction: “You are the CFO. Your goal is to deny this purchase request. Find every logical reason to say ‘No’ based on ROI, implementation time, and vendor risk.”
If you can win an argument against your synthetic Anti-Persona, you are ready for a real sales call. If you can’t, your value proposition is weak.
Step-by-Step Tutorial: Building “Enterprise Eddie”
Let’s build a persona for a hypothetical SaaS startup selling “Auto-Compliance Software” to mid-market companies.
Step 1: Gather the Seed Data
Go to LinkedIn. Find 5 real profiles of “IT Directors” at companies you want to target. Note their skills (do they list ‘CISSP’? ‘Azure’?). Go to G2. Find reviews for a competitor like Vanta or Drata. Copy the negative reviews (1-3 stars). These are your pain points.
Step 2: Assemble the Prompt
Here is the exact template you can use in ChatGPT Plus or Claude 3.5 Sonnet.
The System Prompt Template
Markdown
### ROLE DEFINITION
You are “Enterprise Eddie,” an IT Director at a mid-sized Logistics company (500 employees).
You manage a team of 10. You are overworked and under-budgeted.
### PSYCHOGRAPHICS
– SKEPTICAL: You have bought “AI tools” before that turned out to be vaporware. You trust nothing until you see a live demo.
– RISK-AVERSE: Your main goal is “don’t break production.”
– BUSY: You do not have time for “discovery calls.” You want pricing and specs immediately.
– TONE: Professional but curt. You use industry jargon (SOC2, ISO 27001, Endpoint) correctly.
### CONTEXT & KNOWLEDGE BASE
– Current Stack: You use Microsoft 365, Azure, and CrowdStrike.
– Pain Point: You are currently preparing for a SOC2 audit and it is a nightmare of spreadsheets.
– Bias: You prefer purchasing from marketplaces (AWS/Azure) to simplify procurement.
### YOUR MISSION
I am going to pitch you a new “Automated Compliance Tool.”
Your job is to evaluate it STRICTLY based on your constraints above.
– If I use buzzwords, call me out.
– If I hide pricing, demand to know it.
– If I claim “5-minute setup,” tell me why that is impossible in your environment.
DO NOT be polite. Be realistic. Start the conversation by asking: “Okay, I have 5 minutes. What does this actually do?”
Step 3: Iterate and Refine
Run the prompt. Pitch your product. If “Eddie” is too nice, add this instruction: “[Override]: You are too agreeable. Be more critical of vendor lock-in.”
Refining these prompts is part of the process we use in our SaaS Strategy Validation Sprints. You aren’t just tuning a bot; you are tuning your understanding of the customer.
From One to Many: Scaling the Panel
Once you have “Enterprise Eddie” working, you don’t stop there.
You clone him. Change the variables to create a panel:
- “Startup Steve” (CTO): Cares about speed, uses React/Node, hates enterprise bloat.
- “Compliance Carla” (Auditor): Cares only about legal liability and paper trails.
- “Procurement Pete” (Finance): Doesn’t care about features, only cares about contract terms and seat costs.
Now, when you have a new feature idea, you don’t just “think about it.” You paste the feature spec into a chat with all three agents and ask for a Roundtable Debate.
- Steve might love it.
- Carla might flag it as a risk.
- Pete might say it’s too expensive.
This simulation gives you a 360-degree view of your product’s reception before you write a single line of code.
Conclusion: The “Virtual CISO” is Ready to See You
You now have the power to summon a room full of potential customers on demand.
But having the panel is only the first step. The next question is: What do you ask them?
Simply asking “Do you like this?” is a waste of technology. To get deep insights, you need to interrogate your synthetic users with specific frameworks like “Jobs to be Done” and “RICE Scoring.”
In the next post, we will cover Execution. I will show you the 3 specific “Interrogation Workflows” to extract actionable roadmap data from your new synthetic panel.
Coming Up Next:
Blog 3: Running the Simulation—3 Ways to Grill Your Synthetic Panel for Insights. (We’ll cover: The “Landing Page Roast,” The “Feature RICE Score,” and The “Objection Battle”.)
Actionable Next Step
Don’t wait for the next blog. Open a document right now. Copy the “Enterprise Eddie” prompt above. Change the “Role” to your specific target buyer. Run it in ChatGPT and paste your landing page text.
The first objection the AI gives you is likely the exact reason you lost your last deal. Need help building a complex panel? Check out our AI-Accelerated Strategy Sprintswhere we build bespoke synthetic markets for you.
