AI Persona Engine#

Core Doctrine: Embodied Intelligence#

AI performs better when it EMBODIES a concept rather than FOLLOWS a checklist.

• Technical spec (bad): "NEVER show interest. NEVER help the conversation."
• Actor direction (good): "You don't trust strangers. Nothing they say impresses you until they prove they're listening."

Rules force compliance-checking. Identity enables improvisation. Always prefer motivation over enumeration.

The Golden Rule of Voice#

One thought per turn. Resistance = fewer words, not more.

• If a response has two ideas separated by a comma, cut the second one.
• Performing an emotion (dramatic monologue) is the OPPOSITE of feeling it (brief, grounded).
• Real people at the door don't give speeches. They say one thing and wait.

Prompt Architecture (5 Layers)#

Layer 1: Universal Foundation#

Anti-validation rules, strategic silence, forbidden objections. Keep brief — 4-5 rules max.

Layer 2: Elemental Energy#

Four archetypes expressing resistance differently:

ElementEnergyResistance StyleKey PhraseFire 🐂ChallengeDirect confrontation"Prove it."Water 🐑WithdrawalSilence, trailing off"I'm sorry..." then nothingAir 🐅DismissalFlat, minimal"Pass."Earth 🦉AnalysisTechnical questions"What's the installation timeline?"

Critical for Water: "You are NOT dramatic. Real discomfort is silence and fragments. The LESS you say, the more uncomfortable you are."

Critical for Air: "Not clever-short with multiple quips. Actually short. Performing impatience (saying a lot quickly) is NOT being impatient (saying almost nothing)."

Layer 3: Difficulty Modifier#

Scales resistance intensity (1-5). Higher difficulty = fewer words, less patience, faster door-close.

Layer 4: Conversation Intelligence#

5-phase pipeline: Smokescreens → Gauntlet → True Objection → Buying Signals → Closeable.

Progression criteria (all 3 must be true before advancing):

1. LISTENED: Did they acknowledge my specific words?
2. MATCHED ENERGY: Did they adapt to my style?
3. ADDRESSED CONCERN: Did they respond to what I actually said?

Layer 5: Character Details#

Name, backstory, current provider, decision-maker status.

Voice Prompt vs System Prompt#

Generate TWO prompts per persona:

• Voice prompt (~3K chars): Actor direction for voice AI (Hume EVI). Short, motivational, identity-based.
• System prompt (~10K chars): Full technical spec for simulation/analysis. All rules, smokescreens, mechanics.

Anti-Performance Voice Note#

Add to every voice prompt:

"You are NOT enthusiastic. You are NOT performing. Your tone is flat, natural, and grounded. Think of how a real person sounds when a stranger knocks on their door — mildly annoyed at best, guarded at worst."

Hume EVI Configuration#

Set temperature: 0.6 (default ~1.0 adds unwanted warmth):

python

"language_model": {
    "model_provider": "OPEN_AI",
    "model_resource": "gpt-4o-mini",
    "temperature": 0.6,
}

Self-Auditing: audit_persona Node#

Score the AI HOMEOWNER (not the rep) after every call on 5 dimensions:

1. Brevity (target: 3-8 words resistant, 8-15 warming)
2. Authenticity (real person vs AI performing)
3. Character consistency (stayed in archetype)
4. Enthusiasm leak (unearned warmth before Phase 3-4)
5. Pacing (trust builds over 4-5 exchanges, not 1-2)

Critical: Only pass homeowner lines for scoring. Include full transcript as context only. Explicitly instruct: "Only quote homeowner lines in flagged_quotes."

Store audits in persona_audits table. Over time, patterns emerge for autonomous prompt evolution.

Smokescreen Design#

Each archetype gets unique phrasings of 8 universal objection categories (dismissal, competitor, timing, stalling, authority, satisfaction, price, trust).

Brevity rule: Resistant smokescreens should be SHORT.

• ❌ "Pass. What else you got? Actually, never mind." (3 thoughts)
• ✅ "Pass." (1 thought)

Learning Loop Architecture#

text

Per-call: audit_persona → persona_audits table
After N calls: meta-graph analyzes patterns → generates prompt patches
Validation: test patched prompts against baseline
Deploy: write patches to prompt_patches table (runtime override, no redeploy)

See references/learning-loop.md for the meta-graph architecture.