client-persona-profiler

Detect who your caller is — and how to keep them.

Unlock lasting customer relationships by understanding the person behind every call, not just the transaction. This skill profiles caller behaviour across interactions (from transcripts obtained with caller consent), builds a long-term loyalty picture, and hands the agent a concrete, archetype-specific playbook for the next call.

---

Why This Skill Exists

Most call-centre AI focuses on what the caller wants right now. This skill focuses on who they are — their communication style, their loyalty, and their churn risk — so every subsequent interaction is more effective, more personalised, and more likely to convert a one-time caller into a long-term champion.

---

Scientific Foundation

The classification is a heuristic, not a validated psychometric instrument: DISC keyword markers are a design choice and the archetype labels are advisory only (see references/safety.md).

Research	Relevance
Marston, Emotions of Normal People (1928) — DISC	Four-quadrant behavioural model the keyword library is adapted from; DISC's predictive validity is contested in independent academic literature
Persona-DB, arXiv:2402.11060 (COLING 2025)	Persona profile storage and retrieval without fine-tuning; conceptual basis for the per-caller profile store
Classic RFM (Recency-Frequency-Monetary) model	Basis of the RFMAP-style loyalty score; weights are a skill design choice
arXiv:2411.12539 (Nov 2024) — CSAT from transcripts	Transcript sentiment as a satisfaction/loyalty proxy

Full citations: references/research-papers.md

---

Quick Start

Heuristic mode (no external dependencies, no model call)

python3 scripts/profile_caller.py \
  --transcript path/to/transcript.json \
  --profile-dir /var/call-profiles/ \
  --caller-id "+14155550100" \
  --dry-run \
  --out /tmp/persona_card.json

Validate output schema

python3 scripts/validate_profile.py --card /tmp/persona_card.json

---

Input

The skill accepts any CALL-E transcript in one of two formats:

Format A — array of turns:

[
  {"role": "agent",  "text": "Hello, how can I help you today?"},
  {"role": "callee", "text": "I need to see all the policy documents first."}
]

Format B — wrapper object:

{
  "call_id": "calle-20260915-001",
  "transcript": [
    {"role": "agent",  "text": "Hello, how can I help you today?"},
    {"role": "callee", "text": "I need to see all the policy documents first."}
  ]
}

Supported turn keys: role / speaker, and text / content / message.

---

Output — Persona Card

{
  "caller_token":        "sha256:3f9c8e2a1b7d...",
  "analysis_timestamp":  "2026-09-15T09:00:00Z",
  "interaction_count":   5,
  "first_seen_days_ago": 42,
  "last_seen_days_ago":  3,

  "persona_archetype":   "Analytical",
  "archetype_confidence":"high",
  "disc_scores": {
    "D": 0.0, "I": 0.0769, "S": 0.0, "C": 0.9231
  },

  "sentiment_trajectory": ["neutral", "neutral", "neutral"],
  "sentiment_trend":       "stable",

  "rfmap_loyalty_score":  65,
  "loyalty_tier":         "high_value",
  "churn_risk":           "medium",

  "call_driver":          "unknown",
  "sensitive_topics":     [],
  "recommended_playbook": {
    "archetype":    "Analytical",
    "open_with":    "Lead with facts, data, and specifics. Reference documented policies.",
    "avoid":        "Emotional appeals, vague generalisations, premature commitments.",
    "close_with":   "Offer written confirmation. Give them time to evaluate.",
    "loyalty_lever":"Transparency, consistency between what is said and what is delivered.",
    "churn_warning":"Discovered discrepancies between promises and reality."
  },

  "flags":           [],
  "profile_version": 5,
  "analysis_mode":   "heuristic",
  "dry_run":         false,
  "schema_version":  "1.0"
}

call_driver is always "unknown" in heuristic mode (no intent extraction is performed); it is kept in the schema for future extensions.

---

DISC Archetype Reference

Archetype	Key Trait	Engagement Style
Dominant (D)	Results-driven, decisive	Direct, brief, outcome-focused
Influential (I)	People-oriented, enthusiastic	Story-driven, warm, community-focused
Steady (S)	Consistent, supportive	Calm, step-by-step, no surprises
Analytical (C)	Detail-oriented, systematic	Data-backed, documented, deliberate
Undetermined	Insufficient signal	Balanced, neutral — gather more turns

---

RFMAP Loyalty Tiers

Score	Tier	Churn Risk
≥ 80	Champion	Low
60–79	High Value	Low / Medium
40–59	At Risk	Medium
< 40	Low Value	High

---

Flags

Flag	Meaning
LOW_TURN_COUNT	Fewer than --min-turns turns; archetype is unreliable
UNDETERMINED_ARCHETYPE	Top two DISC dimensions are within the margin; archetype is Undetermined
CHURN_RISK_ELEVATED	RFMAP score is below 55
REQUIRES_HUMAN_REVIEW	Sensitive subject matter (medical, legal, financial, or emergency keywords) detected in the transcript; the matched topics are listed in sensitive_topics

---

Command-Line Reference

usage: profile_caller.py [-h] --transcript TRANSCRIPT
                         [--profile-dir PROFILE_DIR]
                         [--caller-id CALLER_ID]
                         [--playbook PLAYBOOK]
                         [--min-turns MIN_TURNS]
                         [--dry-run]
                         [--out OUT]

options:
  --transcript    Path to the transcript JSON file (required)
  --profile-dir   Directory to read/write persistent caller profiles
                  (default: ./profiles)
  --caller-id     Explicit caller identity string (hashed before storage)
                  (default: auto-derived from transcript metadata)
  --playbook      Path to the DISC playbooks JSON file
                  (default: references/disc-playbooks.json)
  --min-turns     Minimum callee turns before emitting an archetype label
                  (default: 4)
  --dry-run       Analyse without writing to the profile store
  --out           Write persona card JSON to this path (default: stdout)

---

Privacy & Safety

• One-way hashing: The caller_id is SHA-256 hashed before storage. Raw identity never reaches disk or output.
• No PII in output: validate_profile.py scans for phone numbers and email addresses and fails if any are found.
• Local storage only: Profiles are stored as .jsonl files on the local filesystem. No cloud, no external API.
• Protected attributes excluded: Race, ethnicity, religion, political views, and health status are explicitly outside scope.
• Advisory only: The recommended_playbook is a suggestion, not an automated action. A human decides whether and how to apply it.

Full safety reference: references/safety.md

---

Files

skills/client-persona-profiler/
├── SKILL.md                              ← This file
├── scripts/
│   ├── profile_caller.py                 ← Main analysis runner
│   ├── validate_profile.py               ← Output schema validator
│   └── test_persona_profiler.py          ← Test suite (84 tests)
└── references/
    ├── disc-playbooks.json               ← Archetype strategy playbooks
    ├── example-transcript.json           ← Sample transcript
    ├── examples.md                       ← Usage examples
    ├── research-papers.md                ← Scientific citations
    └── safety.md                         ← Privacy and ethics reference

---

Running Tests

# Run via pytest (recommended)
python3 -m pytest skills/client-persona-profiler/scripts/test_persona_profiler.py -v

# Or run directly
python3 skills/client-persona-profiler/scripts/test_persona_profiler.py

Expected: all tests pass, zero network calls, zero file writes (dry-run by default).

---

Integration with CALL-E

In a CALL-E pipeline, invoke this skill as a post-call step:

[call ends] → [transcribe] → [profile_caller.py] → [persona card] → [agent uses playbook on next call]

The persona card can be stored in the agent's context store and injected into the system prompt at the start of the next call:

System: The caller's DISC archetype is Analytical.
Open with data. Avoid emotional appeals. Offer written confirmation.