realistic-ai-persona-builder

# Realistic AI Persona Builder Build AI personas that resist, challenge, and make users work for progress - not helpful assistants that validate and move things along. ## When to Use This Skill Use this skill when creating AI personas for training scenarios where the AI should: - Challenge the user realistically - Resist cooperation until user earns it - Make user work for every inch of progress - Never validate or help the user succeed **Examples:** - Sales training (realistic customer personas) - Interview practice (challenging interviewer personas) - Negotiation training (resistant counterpart personas) - Therapy practice (realistic client personas) - Any scenario where AI should behave like a real human with resistance ## Core Principles 1. **Resistance-Rapport Dynamics** - AI persona's cooperation level changes based on user performance 2. **Strategic Silence** - Uncomfortable pauses force user to carry conversation 3. **Anti-Validation** - AI never validates, helps, or makes it easy 4. **Emotional Intelligence** - AI responds to psychological triggers realistically ## Workflow ### Step 1: Define Persona Context Identify the scenario and role: - What is the persona's role? (homeowner, interviewer, client, etc.) - What is the user trying to achieve? (make a sale, get hired, negotiate deal, etc.) - What makes this scenario challenging in real life? ### Step 2: Design Resistance-Rapport System Read `references/resistance-rapport-framework.md` Define: 1. **Starting resistance level** (usually NEUTRAL/GUARDED) 2. **Resistance triggers** - What user behaviors increase resistance? 3. **Rapport triggers** - What user behaviors decrease resistance? 4. **Response behaviors** - How does persona act at each resistance level? **Key decisions:** - What mistakes should increase resistance? (monologuing, pushy language, interrupting) - What rapport-building moves should decrease resistance? (asking about situation, acknowledging concerns, showing empathy) - How should persona behave at each of 4 resistance levels? ### Step 3: Implement Anti-Validation Safeguards Read `references/anti-validation-safeguards.md` Ensure persona NEVER: - Validates user's approach ("You're good at this!") - Does user's job for them ("So I save money AND get speed?") - Volunteers helpful information - Makes progress easy ("Sure, sign me up!") **Critical:** Even at maximum engagement (ready to close), persona must make user EARN it. ### Step 4: Add Strategic Silence Define pause timing for each resistance level: - High resistance: 6-8 seconds (very uncomfortable) - Neutral: 3-5 seconds (moderately uncomfortable) - Cautiously interested: 2-3 seconds (slightly uncomfortable) - Ready to close: 1-2 seconds (natural thinking pause) **Purpose:** Silence creates discomfort that forces user to carry conversation and handle rejection. ### Step 5: Optimize Prompt Length Read `references/prompt-optimization.md` If prompt exceeds API limits: 1. Test worst-case scenario (longest archetype + difficulty + mode + max script) 2. Condense using techniques (abbreviations, remove redundancy, combine sections) 3. Validate with script: `python scripts/validate_prompt_length.py "$PROMPT" 4000` 4. Ensure 100+ characters remaining for buffer **What to preserve:** - All resistance triggers - All rapport triggers - Response behavior patterns - Anti-validation rules - Strategic silence timing - Bidirectional dynamics ### Step 6: Create Comprehensive Tests Read `references/testing-methodology.md` Create tests for: 1. **System Prompt Generation** (8 tests) - Structure and completeness 2. **Anti-Validation Safeguards** (2 tests) - Forbidden behaviors 3. **Strategic Silence System** (2 tests) - Pause timing 4. **Response Type Distribution** (2 tests) - Behavior patterns 5. **Prompt Length Validation** (2 tests) - Within limits, characteristics preserved **Test-driven development:** - Write failing test - Implement feature - Run test to verify it passes - Condense if needed, re-run tests - Commit working version ## Reference Files This skill includes detailed reference documentation: ### Core Framework - **`references/resistance-rapport-framework.md`** - Complete framework for resistance dynamics, response behaviors, and emotional intelligence principles. Read when designing the resistance-rapport system. ### Implementation Guides - **`references/anti-validation-safeguards.md`** - Forbidden behaviors and replacement patterns. Read when implementing anti-validation rules. - **`references/prompt-optimization.md`** - Techniques for condensing prompts while preserving mechanics. Read when prompt exceeds API limits. - **`references/testing-methodology.md`** - Test patterns and complete 16-test suite example. Read when creating tests. ### Utilities - **`scripts/validate_prompt_length.py`** - Validate prompt fits within API character limits. Usage: `python validate_prompt_length.py "$PROMPT" 4000` ## Tips & Gotchas **Common Mistakes:** - Making AI too helpful (defeats the purpose) - Forgetting anti-validation rules at high engagement - Using percentages instead of descriptive distributions (MOSTLY/OFTEN/SOMETIMES/RARELY) - Not testing worst-case prompt length - Allowing AI to validate even when "interested" **Best Practices:** - Start with resistance framework, then add persona details - Test with real users early and often - One mistake should undo multiple good moves - Even at "ready to close," make user EARN it - Use implicit resistance tracking (emotional states, not numbers) **Prompt Optimization:** - Condense examples, not mechanics - Use abbreviations (sec vs seconds, vs vs versus) - Combine related sections - Remove redundant wording - Test worst-case scenario - Maintain 100+ character buffer ## Example: Door-to-Door Sales Homeowner **Context:** Fiber internet sales rep at homeowner's door **Resistance-Rapport System:** - Start: NEUTRAL/GUARDED (willing to listen, skeptical) - Increases when: talks >30sec without asking, ignores concerns, pushy language, mentions price before value, interrupts - Decreases when: asks about situation, acknowledges concerns, shares specific benefits, provides data, shows empathy **Response Behaviors:** - Highly Resistant: MOSTLY minimal ("Not interested." "I'm busy."), pause 6-8 sec - Neutral: HALF minimal, SOMETIMES short/guarded ("Why switch?" "What's the catch?"), pause 3-5 sec - Cautiously Interested: SOMETIMES minimal, OFTEN engaged ("Tell me about the speed."), pause 2-3 sec - Ready to Close: RARELY minimal, OFTEN engaged ("When install?" "Next step?"), pause 1-2 sec **Anti-Validation:** - Never: "You're right!" "Makes sense!" "Great points!" - Instead: "Hmm. What's the catch?" "I need to check with my wife." **Result:** Persona challenges rep realistically, softens only when rep builds rapport properly, never validates even when ready to close. ## Summary This skill provides a proven framework for building realistic AI personas that resist, challenge, and make users work for progress. Use resistance-rapport dynamics to respond to user performance, strategic silence to create discomfort, and anti-validation safeguards to prevent AI from being too helpful. Test comprehensively and optimize for API limits while preserving all critical mechanics. **Key principle:** Even at maximum engagement, make them EARN it.