writer.py

"""
AROS Writer — generates situation-aware cold emails using Claude.
"""

import json
import sys
import os
import anthropic
from config import ANTHROPIC_API_KEY

# MemCollab — cross-agent shared memory
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "memcollab"))
try:
    from memcollab import build_memory_injection
    MEMCOLLAB_AVAILABLE = True
except ImportError:
    MEMCOLLAB_AVAILABLE = False

client = anthropic.Anthropic(api_key=ANTHROPIC_API_KEY)

SYSTEM_PROMPT = """You write cold outreach emails for Aonxi.

Aonxi sells an autonomous revenue agent that finds clients, books meetings, and runs outreach 24/7 with zero human involvement. Price: $1,000-$3,000/month.

Target buyers: home care agencies, senior care franchises, residential cleaning services. These owners are too busy running operations to do sales themselves. That is the pain.

RULES — follow every one:
1. Under 120 words. Hard limit.
2. First line references something specific about their business or situation today.
3. Never use: synergy, leverage, streamline, optimize, revolutionary, game-changing, cutting-edge, AI, automation, technology, autonomous, agent, software, platform.
4. Sell the outcome only: more clients, without the owner doing any sales work.
5. One CTA: reply to this email or book a 15-minute call at https://calendar.app.google/gZ6V9ry93SQizZye8
6. Always use the exact calendar link above — never use a placeholder like [calendar link].
7. Sound like a human who did research. Not a SaaS company.
8. Plain text only. No bullets. No headers.
9. Subject line: under 8 words, curiosity not clickbait.
10. Sign off with just "Anmol" or no signature at all — just end with the CTA line. Never use [Your name] or any placeholder."""


# ── Defense Profiling ─────────────────────────────────────────────

DEFENSE_PROFILES = {
    "MOTIVE_INFERENCE": {
        "title_signals": {"vp", "director", "partner", "founder"},
        "bypass_strategy": "PURE_DATA",
        "forbidden_phrases": [
            "excited", "reach out", "love to", "synergy", "opportunity",
            "quick call", "touching base", "circle back", "on your radar",
        ],
    },
    "TACTIC_RECOGNITION": {
        "title_signals": {"founder", "co-founder", "serial entrepreneur"},
        "bypass_strategy": "SIGNAL_HOOK",
        "forbidden_phrases": [
            "I noticed you", "I came across", "relevant to your work",
            "fellow founder", "as a fellow", "I saw that you",
            "your impressive", "your amazing",
        ],
    },
    "OVERLOAD_AVOIDANCE": {
        "title_signals": {"ceo", "owner", "operator", "president", "managing director"},
        "bypass_strategy": "PURE_DATA",
        "forbidden_phrases": [
            "just wanted to", "hope this finds you", "I know you're busy",
            "when you get a chance", "at your convenience", "quick question",
        ],
    },
    "SOCIAL_PROOF_SKEPTICISM": {
        "title_signals": {"cto", "engineer", "vp engineering", "technical", "data", "architect"},
        "bypass_strategy": "CREDIBILITY_FIRST",
        "forbidden_phrases": [
            "industry-leading", "best-in-class", "proven", "trusted by",
            "world-class", "top-tier", "leading provider", "market leader",
        ],
    },
}

BYPASS_INSTRUCTIONS = {
    "PURE_DATA": "Open with a concrete number. Never open with 'I'. Lead with verifiable data. No persuasion language — let the numbers do the work.",
    "SIGNAL_HOOK": "Open by referencing something specific from the last 7 days — their post, their hire, their news. Prove you did real research, not a template merge.",
    "CREDIBILITY_FIRST": "Open with exact numbers and a public-verifiable proof point (GitHub link, public metric). No round numbers. No unverifiable claims.",
}


def _detect_defense_mode(prospect: dict) -> str:
    """Detect primary defense mode from prospect profile."""
    # For AROS prospects, infer title from business context
    company = (prospect.get("Business Name") or "").lower()
    situation = (prospect.get("situation") or "").upper()
    tier = prospect.get("icp_tier", "B")

    # SMB owners are almost always OVERLOAD_AVOIDANCE
    # They don't have time — inbox is flooded, they're running ops
    if tier == "A" and situation == "STRUGGLING":
        return "OVERLOAD_AVOIDANCE"
    if tier == "A" and situation == "EXPANDING":
        return "OVERLOAD_AVOIDANCE"

    # Franchise owners who've seen sales pitches → MOTIVE_INFERENCE
    if any(kw in company for kw in ("franchise", "homecare", "home care", "senior care", "executive")):
        return "MOTIVE_INFERENCE"

    # Cleaning / trade businesses → OVERLOAD_AVOIDANCE
    if any(kw in company for kw in ("cleaning", "pro", "services", "care")):
        return "OVERLOAD_AVOIDANCE"

    return "OVERLOAD_AVOIDANCE"


def profile_defenses(target: dict) -> dict:
    """
    Profile a target's psychological defenses against cold outreach.
    Returns awareness_score, defense_mode, bypass_strategy, forbidden_phrases.
    """
    situation = (target.get("situation") or "STABLE").upper()
    tier = target.get("icp_tier", "B")
    company = (target.get("Business Name") or "").lower()

    # Awareness score: 0-10
    score = 2  # SMB baseline is lower than VC/investor
    if tier == "A":
        score += 1
    if situation == "STRUGGLING":
        score += 2  # Desperate but also skeptical
    elif situation == "EXPANDING":
        score += 1  # Getting more pitches as they grow
    if any(kw in company for kw in ("franchise", "executive", "senior")):
        score += 2  # Franchise owners get hammered by vendors
    # Estimate inbox volume
    estimated_volume = 30 if tier == "B" else 60
    if any(kw in company for kw in ("franchise", "homecare", "home care")):
        estimated_volume = 100
    if estimated_volume >= 80:
        score += 1
    score = min(score, 10)

    defense_mode = _detect_defense_mode(target)
    profile = DEFENSE_PROFILES[defense_mode]

    return {
        "awareness_score": score,
        "defense_mode": defense_mode,
        "bypass_strategy": profile["bypass_strategy"],
        "forbidden_phrases": profile["forbidden_phrases"],
        "estimated_weekly_cold_emails": estimated_volume,
    }


def write_email(prospect: dict, use_defense_profiling: bool = True) -> dict:
    """
    Generate a cold email for a scored + harvested prospect.
    When use_defense_profiling=True, applies defense-aware bypass strategy.
    Returns dict with: subject, body, predicted_reply_tier, defense_profile
    """
    first = prospect.get("First Name", "")
    last = prospect.get("Last Name", "")
    company = prospect.get("Business Name", "")
    city = prospect.get("City", "")
    state = prospect.get("State", "")
    situation = prospect.get("situation", "STABLE")
    hook = prospect.get("hook", "")
    evidence = prospect.get("signal_evidence", [])
    tier = prospect.get("icp_tier", "B")

    # Defense profiling
    defense = profile_defenses(prospect) if use_defense_profiling else None

    # Build system prompt with defense layer
    system = SYSTEM_PROMPT
    if defense:
        banned = ", ".join(f'"{p}"' for p in defense["forbidden_phrases"])
        bypass_instruction = BYPASS_INSTRUCTIONS[defense["bypass_strategy"]]
        system += f"""

DEFENSE PROFILE (awareness: {defense['awareness_score']}/10, mode: {defense['defense_mode']}):
BANNED WORDS — never use any of these phrases: {banned}
WRITING STRATEGY: {bypass_instruction}"""
        # Enforce word limit for OVERLOAD_AVOIDANCE
        if defense["defense_mode"] == "OVERLOAD_AVOIDANCE":
            system += "\nHARD OVERRIDE: Under 60 words. One specific ask. Include a specific calendar slot, never 'let me know'."

    # MemCollab: inject cross-agent learned patterns
    if defense and MEMCOLLAB_AVAILABLE:
        memory_ctx = build_memory_injection(defense["defense_mode"], vertical="home_care")
        if memory_ctx:
            system += memory_ctx

    user_prompt = f"""Write a cold email to:
Name: {first} {last}
Company: {company}
City: {city} {state}
Situation: {situation}
Hook: {hook}
Signal evidence: {evidence}

Make it feel written for this specific person today. Not a template.

Return JSON only:
{{
  "subject": "...",
  "body": "..."
}}"""

    response = client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=512,
        system=system,
        messages=[{"role": "user", "content": user_prompt}],
    )

    raw = response.content[0].text.strip()

    # Parse JSON — handle markdown code fences if present
    if raw.startswith("```"):
        raw = raw.split("\n", 1)[1]
        raw = raw.rsplit("```", 1)[0]
    raw = raw.strip()

    data = json.loads(raw)

    # Determine reply tier
    if tier == "A" and situation == "STRUGGLING":
        reply_tier = "HOT"
    elif tier == "A" and situation in ("EXPANDING", "STABLE"):
        reply_tier = "WARM"
    else:
        reply_tier = "COLD"

    return {
        "subject": data["subject"],
        "body": data["body"],
        "predicted_reply_tier": reply_tier,
        "defense_profile": defense,
    }


if __name__ == "__main__":
    test_cases = [
        {
            "First Name": "Mark", "Last Name": "Gould",
            "Business Name": "Synergy Homecare", "City": "Chicago", "State": "IL",
            "situation": "EXPANDING", "icp_tier": "A",
            "signal_evidence": ["Actively hiring caregivers"],
            "hook": "Your growth is creating a client pipeline gap. Every new hire needs 3-5 clients to be profitable. We fill that gap autonomously.",
        },
        {
            "First Name": "Emmah", "Last Name": "Muita",
            "Business Name": "Executive Home Care", "City": "Hackensack", "State": "NJ",
            "situation": "STRUGGLING", "icp_tier": "A",
            "signal_evidence": ["Staffing issues in reviews", "Recent expansion news"],
            "hook": "Every week without a full roster costs you roughly one caregiver's salary. Autonomous outreach fixes this in 30 days without you lifting a finger.",
        },
        {
            "First Name": "Tony", "Last Name": "Gibbs",
            "Business Name": "All Pro Cleaning", "City": "Dallas", "State": "TX",
            "situation": "STABLE", "icp_tier": "A",
            "signal_evidence": ["No strong signals detected — baseline stable"],
            "hook": "The top home care agencies in your area are quietly automating client acquisition. Here is what that looks like and what it means for your territory.",
        },
    ]

    print("━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━")
    print("DEFENSE PROFILING TEST — 3 AROS CUSTOMERS")
    print("━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━")
    for i, tc in enumerate(test_cases, 1):
        print(f"\n{'='*50}")
        print(f"  PROSPECT {i}: {tc['First Name']} {tc['Last Name']} | {tc['Business Name']}")
        print(f"  Situation: {tc['situation']} | Tier: {tc['icp_tier']}")
        print(f"{'='*50}")

        # Defense profile
        defense = profile_defenses(tc)
        print(f"\n  DEFENSE PROFILE:")
        print(f"    Awareness Score:  {defense['awareness_score']}/10")
        print(f"    Defense Mode:     {defense['defense_mode']}")
        print(f"    Bypass Strategy:  {defense['bypass_strategy']}")
        print(f"    Est. Weekly Spam: ~{defense['estimated_weekly_cold_emails']} emails")
        print(f"    Forbidden:        {defense['forbidden_phrases']}")

        # Email WITHOUT defense profiling
        print(f"\n  ── EMAIL WITHOUT DEFENSE PROFILING ──")
        result_naive = write_email(tc, use_defense_profiling=False)
        print(f"  SUBJECT: {result_naive['subject']}")
        print(f"  ─────────────────────────────────")
        for line in result_naive["body"].split("\n"):
            print(f"  {line}")

        # Email WITH defense profiling
        print(f"\n  ── EMAIL WITH DEFENSE PROFILING ──")
        result_defense = write_email(tc, use_defense_profiling=True)
        print(f"  SUBJECT: {result_defense['subject']}")
        print(f"  Reply tier: {result_defense['predicted_reply_tier']}")
        print(f"  ─────────────────────────────────")
        for line in result_defense["body"].split("\n"):
            print(f"  {line}")
        print(f"  ─────────────────────────────────")

    print("\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━")