ClientCast Generator

Towards a Client-Centered Assessment of LLM Therapists by Client Simulation

Paper: arXiv

Extracts a structured psychological profile from a real therapy conversation, including Big Five personality traits and PHQ-9 / GAD-7 / OQ-45 symptom estimates.

Overview

Property	Value
Key	clientCast
Type	LLM-based
Output	ClientCast character files

Key Features

• Basic profile extraction: Infers demographics, presenting problem, emotional style, and resistance level from conversation text
• Big Five personality traits: Estimates Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism with percentage scores
• Clinical scale estimation: Scores PHQ-9, GAD-7, and OQ-45 items from the conversation

How It Works

generate_character() runs three sequential LLM calls:

1. Basic profile — extracts name, gender, age, occupation, topic, situation, emotion, resistance, and emotional fluctuation
2. Big Five — estimates each trait as a percentage score with explanation
3. Symptoms — scores individual items from PHQ-9, GAD-7, and OQ-45

The result is saved to output_dir.

Usage

from patienthub.generators import get_generator

generator = get_generator(agent_name="clientCast", lang="en")
generator.generate_character()

Configuration

Parameter	Type	Default	Description
prompt_path	string	data/prompts/generator/clientCast.yaml	Path to prompt file
input_dir	string	data/resources/ClientCast/human_data.json	Path to input conversation JSON
symptoms_dir	string	data/resources/ClientCast/symptoms.json	Path to symptom item definitions
output_dir	string	data/characters/ClientCast.json	Path where the generated character is saved
data_idx	int	0	Index of the conversation to use from input file
model_type	string	"OPENAI"	Model provider key
model_name	string	"gpt-4o"	Model identifier
temperature	float	0.7	Sampling temperature
max_tokens	int	8192	Max response tokens
max_retries	int	3	API retry attempts

Input Data Format

The input file is a JSON array of conversation objects:

[
  {
    "messages": [
      {"role": "Therapist", "content": "What brings you here today?"},
      {"role": "Client", "content": "I've been feeling really low lately..."}
    ]
  }
]

data_idx selects which conversation to process.

Output Format

{
  "basic_profile": {
    "name": "Not Specified",
    "gender": "Female",
    "age": 34,
    "occupation": "Teacher",
    "topic": "smoking cessation – wants to quit but struggles with cravings.",
    "situation": "The client feels overwhelmed by work stress and persistent low mood.",
    "emotion": "The client feels anxious and guilty about disappointing others.",
    "resistance": "Low – the client is open to suggestions."
  },
  "big_five": {
    "Openness": {"score_percent": 62, "explanation": "..."},
    "Conscientiousness": {"score_percent": 55, "explanation": "..."}
  },
  "symptoms": {
    "PHQ-9": {"1": {"identified": true, "severity_level": 2, "severity_label": "More than half the days", "explanation": "..."}},
    "GAD-7": {"1": {"identified": true, "severity_level": 1, "explanation": "..."}},
    "OQ-45": {"1": {"identified": false, "explanation": "..."}}
  }
}

Use Cases

• Creating ClientCast character files from existing therapy transcripts
• Building assessment datasets from real or simulated conversations