From generation to validation

Parts 1 and 2 covered techniques for getting AI to produce good outputs: clear instructions, context, examples, and step-by-step reasoning.

This section focuses on validation and critical analysis—techniques that help you verify quality, find errors, and consider multiple perspectives before making decisions.

Technique 6: Role prompting

Role prompting assigns the AI a specific perspective, expertise, or persona. This shapes how it approaches problems and what factors it prioritizes.

The principle

Different roles bring different perspectives. A CFO evaluates decisions differently than a CTO. A skeptical customer sees products differently than an enthusiastic early adopter.

By assigning roles, you can:

  • •Access specialized reasoning patterns
  • •Simulate stakeholder perspectives before real meetings
  • •Identify concerns you might otherwise miss
  • •Generate diverse viewpoints on the same problem

What this looks like

Without role:

"What do you think about launching this feature?"

With role:

"You are a senior product manager with 10 years of experience at consumer tech companies. You've seen many feature launches succeed and fail. Evaluate this feature proposal, focusing on user adoption risks and go-to-market considerations."

The role shapes both what the AI considers important and how it communicates.

Practical examples

Everyday use:

"You are a professional editor reviewing my cover letter. Identify weak phrases, unclear statements, and areas where I could be more specific. Be direct—I want honest feedback, not encouragement."

Product Manager:

"Analyze this feature from three different user perspectives:

  1. •As a power user who uses our product daily for 2+ hours
  2. •As a new user in their first week
  3. •As an administrator managing a team of 50

What would each user think about this feature? What concerns would they have?"

Founder:

"You are a venture capitalist evaluating our pitch deck. You've seen 1,000 decks this year and funded 10 companies. What questions would you ask? What concerns would make you pass? What would need to be true for you to invest?"

How companies use this: Multi-perspective analysis

Companies building AI for research, analysis, and decision-support use role prompting to generate comprehensive viewpoints.

Legal AI (like Harvey or CoCounsel):

"Analyze this contract from two perspectives:

As the party signing this agreement:
Identify clauses that create risk or unfavorable obligations. Flag anything unusual or concerning. Suggest modifications.

As opposing counsel who drafted this:
Explain why each clause was likely included. What is the other party trying to protect or achieve?

Contract: [Document text]"

Market research AI (like Qualtrics or SurveyMonkey AI features):

"Analyze this product concept from different customer segments:

As a budget-conscious consumer:
What matters most? What would make you hesitate to purchase?

As a premium buyer who values quality over price:
What signals quality to you? What would make this feel worth the investment?

As a business buyer making decisions for a team:
What would you need to justify this purchase to your manager?

Product concept: [Description]"

Sales enablement AI (like Gong or Clari):

"Analyze this sales call transcript from multiple angles:

As the customer:
What were my real concerns? Did the salesperson address them? What would have convinced me?

As a sales coach:
What did the rep do well? What opportunities did they miss? How could they improve?

As a competitor:
What weaknesses in our offering were revealed? How would I position against us?

Transcript: [Call text]"

Technique 7: Critique prompting

Critique prompting explicitly asks the AI to find problems, errors, or gaps. Instead of asking "Is this good?", you ask "What's wrong with this?"

The principle

AI systems, like humans, can be agreeable by default. Asking for validation often produces validation. Asking for criticism produces more useful feedback.

Critique prompting helps you:

  • •Find errors before they become problems
  • •Identify gaps in logic or missing considerations
  • •Stress-test ideas before committing resources
  • •Improve work through specific, actionable feedback

What this looks like

Seeking validation (less useful):

"Does this business plan look good?"

Seeking critique (more useful):

"You are a critical analyst. Find every weakness in this business plan. What assumptions are we making that might be wrong? What risks are we underestimating? What would cause this to fail? Be thorough—I'd rather hear the problems now than after we've invested."

Practical examples

Everyday use:

"Critique this email I'm about to send to my team. Point out anything that could be misunderstood, comes across as passive-aggressive, or is unclear. Suggest specific improvements."

Product Manager:

"Here's our product roadmap for Q1. Challenge every item:

  • •What's missing that should be here?
  • •What's included that shouldn't be prioritized?
  • •What dependencies or risks have we overlooked?
  • •Where are we being overly optimistic about timelines?

Be specific—vague concerns aren't helpful."

Founder:

"We're planning to raise a Series A in 6 months. Here's our current position. Tell me everything that could prevent us from raising successfully:

  • •Metrics: $80K MRR, 15% month-over-month growth, 8% churn
  • •Team: 5 full-time, 2 contractors
  • •Market: Project management for remote teams

What would make an investor say no? What should we fix before fundraising?"

How companies use this: Quality assurance and risk assessment

Companies use critique prompting to build quality checks into AI systems.

Code review AI (like GitHub Copilot or Sourcegraph):

"Review this code for problems:

  1. •Security vulnerabilities (SQL injection, XSS, authentication issues)
  2. •Performance problems (inefficient queries, memory leaks, unnecessary loops)
  3. •Maintainability issues (unclear naming, missing documentation, complex logic)
  4. •Edge cases not handled

For each issue found, explain the problem, the risk level, and how to fix it.

Code: [Code block]"

Content moderation AI (like Perspective API or platform-specific tools):

"Analyze this content for policy violations:

  1. •Harassment or bullying (targeting individuals, threatening language)
  2. •Misinformation (false claims, misleading context)
  3. •Spam signals (promotional language, repetitive posting patterns)
  4. •Harmful content (self-harm, dangerous activities)

For each potential issue, rate severity (low/medium/high) and confidence level. Explain your reasoning.

Content: [User post]"

Financial risk AI (like Kensho or Ayasdi):

"Analyze this loan application for risk factors:

  1. •Credit risk indicators
  2. •Income stability concerns
  3. •Debt-to-income ratio assessment
  4. •Collateral adequacy
  5. •Market/industry risk factors

Flag anything concerning. For each flag, explain the risk, severity, and what additional information would help assess it.

Application: [Applicant data]"

Technique 8: Self-consistency and verification

Self-consistency involves asking the AI to verify its own work or approach the same problem from multiple angles to check for consistent answers.

The principle

For important decisions, a single answer isn't enough. Verification catches errors and builds confidence. You can:

  • •Ask the AI to check its own work
  • •Request the same analysis using different methods
  • •Have it argue the opposite position to test robustness
  • •Ask it to identify what would change its conclusion

What this looks like

Without verification:

"Calculate the ROI of this project."

With verification:

"Calculate the ROI of this project. Show your work.

Then verify your calculation:

  1. •Check each number—are the inputs accurate?
  2. •Recalculate using a different method if possible
  3. •Identify what assumptions you made
  4. •State how confident you are in the result (high/medium/low) and why"

Practical examples

Everyday use:

"I calculated that I should refinance my mortgage to save money. Here are the numbers: [details].

Verify this analysis:

  1. •Check if my math is correct
  2. •Identify costs I might have forgotten
  3. •Point out assumptions that might be wrong
  4. •Tell me what scenarios would make refinancing a bad decision"

Product Manager:

"I've concluded we should prioritize Feature A over Feature B based on this analysis: [analysis].

Now argue the opposite. Make the strongest possible case for Feature B. What am I missing or underweighting? After presenting that argument, tell me if you still agree with the original recommendation or if the counterargument changes things."

Founder:

"Based on our data, I believe we should expand to the enterprise market. Here's my reasoning: [reasoning].

Stress test this conclusion:

  1. •What evidence would contradict this decision?
  2. •What would need to be true for this to be wrong?
  3. •What's the strongest argument against enterprise expansion?
  4. •Rate your confidence in the original recommendation after considering these challenges."

How companies use this: High-stakes verification

For critical applications, companies build verification loops into their AI systems.

Medical AI (like Tempus or PathAI):

"Analyze this diagnostic image. Provide your assessment.

Then perform verification:

  1. •What findings support your conclusion?
  2. •What findings might suggest alternative diagnoses?
  3. •What additional tests would increase diagnostic certainty?
  4. •Rate confidence: high (>90%), medium (70-90%), low (<70%)
  5. •If confidence is not high, recommend physician review.

Image: [Medical scan]"

Fraud detection AI (like Stripe Radar or Sift):

"Evaluate this transaction for fraud risk.

First analysis: Standard risk assessment based on transaction patterns.

Verification layer:

  1. •What legitimate explanations exist for any unusual patterns?
  2. •What additional signals would confirm or rule out fraud?
  3. •Compare to similar transactions that were later confirmed as fraud/legitimate
  4. •Final recommendation: Approve / Review / Block
  5. •Confidence level and reasoning

Transaction: [Transaction data]"

Autonomous systems (like Waymo or Tesla Autopilot decision-making):

"Assess the current driving situation.

Primary analysis: Identify obstacles, predict movements, determine safe path.

Verification:

  1. •Cross-check obstacle detection with multiple sensor inputs
  2. •Consider alternative interpretations of sensor data
  3. •Identify edge cases or unusual conditions
  4. •Confidence level for chosen action
  5. •If confidence below threshold, escalate to human or take conservative action

Sensor data: [Input streams]"

Combining validation techniques

These techniques work together for comprehensive analysis.

Combined Example: Product Decision Review

"We're deciding whether to build an AI feature for our product. Here's the proposal: [details].

Role Analysis: Evaluate this proposal from three perspectives:

  • •Engineering Lead: Technical feasibility and resource concerns
  • •CFO: Financial implications and ROI
  • •Customer Success: Adoption and support implications

Critique: For each perspective, focus on problems, risks, and concerns—not benefits.

Verification: After completing the analysis, identify:

  • •Which concerns are most serious?
  • •What would change the conclusion?
  • •Overall recommendation: Proceed / Modify / Reconsider
  • •Confidence level and key uncertainties"

Key takeaways

  • •Role prompting accesses different perspectives. Assign specific roles to get specialized viewpoints, simulate stakeholder reactions, or generate diverse analysis.
  • •Critique prompting finds problems proactively. Explicitly ask for criticism, weaknesses, and risks. What you don't know can hurt you.
  • •Verification builds confidence in important decisions. Check work, argue the opposite, and identify what would change conclusions.
  • •Companies use these techniques for quality assurance. From legal review to fraud detection to medical diagnosis, validation techniques make AI systems more reliable.

What's next

Part 4 addresses systematic improvement: how to test prompts effectively, compare results across different AI models, and build repeatable processes for prompt optimization.

This is Part 3 of a four-part series on prompt engineering techniques.