conjoint-study

name	conjoint-study
description	Design, execute, and analyze conjoint analysis studies - guides through attribute/level definition, experimental design, EDSL survey generation, and part-worth utility estimation
allowed-tools	Read, Write, Glob, Grep, Skill, AskUserQuestion, Bash(python3:*)
user-invocable	true
arguments	research_question

Conjoint Study

Takes a research question and guides the user through designing a conjoint analysis study, generating balanced experimental designs, creating EDSL project files, running the study, and analyzing results with part-worth utilities.

Usage

/conjoint-study What features matter most when choosing a streaming service?
/conjoint-study How do job seekers trade off salary, remote work, and company size?
/conjoint-study What attributes drive consumer preference for electric vehicles?

Workflow

Path Discovery: This skill bundles a helpers.py script. Before running it, use Glob("**/conjoint-study/helpers.py") to locate it. Store the result in a variable and use it for all subsequent calls.

Phase 0: Setup

1. Locate helpers.py using Glob("**/conjoint-study/helpers.py").

2. Create the project directory using the helper:

python3 <helpers_path> setup-dir "<research_question>"

This prints two lines: the directory name (e.g. 2026-02-10_streaming-service-preferences) and either NEW or EXISTS.

3. Check for existing design. If EXISTS, use AskUserQuestion:

Question: "A conjoint design already exists at <dir_name>/conjoint_design.md. What would you like to do?"
Header: "Existing file"
Options:
  1. "Start fresh" - "Overwrite the existing design with a new one"
  2. "Modify existing" - "Read the current design and revise it"
  3. "Cancel" - "Keep the existing file unchanged"

All output files are written into <dir_name>/.

Phase 1: Elicit Study Design

Conduct 6 AskUserQuestion interactions to gather the full study specification.

1a. Research Question

If the user's input is vague, clarify:

Question: "Could you describe your conjoint study in more detail? Specifically: What product/service are you studying? What decision are respondents making? What trade-offs are you interested in?"
Header: "Clarify"

Parse from the input:

• Product/service category being evaluated
• Decision context (purchase, subscription, job choice, etc.)
• Key trade-offs the researcher wants to measure

1b. Attributes and Levels

Ask the user to define attributes and their levels:

Question: "Please list the attributes and levels for your conjoint study. Format each as 'Attribute: Level1, Level2, Level3'. For example:\n\nPrice: $9.99, $14.99, $19.99\nContent Library: Small, Medium, Large\nAd Experience: No ads, Limited ads, Full ads\n\nGuidelines:\n- Use 3-8 attributes\n- Use 2-5 levels per attribute\n- Attributes should be actionable and independent\n- Levels should span a realistic range\n- Avoid 'must-have' attributes where one level always dominates"
Header: "Attributes"

After receiving the response, parse the attributes and levels. Validate:

• 3-8 attributes (warn if outside range)
• 2-5 levels per attribute (warn if outside range)

Present a formatted confirmation table:

Question: "Here are the attributes and levels I've parsed:\n\n| Attribute | Levels |\n|-----------|--------|\n| Price | $9.99, $14.99, $19.99 |\n| ... | ... |\n\nTotal unique profiles: <N>\n\nDoes this look correct?"
Header: "Confirm attrs"
Options:
  1. "Looks good" - "Proceed with these attributes and levels"
  2. "Make changes" - "Edit the attributes or levels"

1c. Conjoint Method

Question: "Which conjoint method would you like to use?"
Header: "Method"
Options:
  1. "Choice-Based Conjoint (CBC) (Recommended)" - "Respondents choose their preferred option from a set of profiles. Most widely used method."
  2. "Ranking-Based" - "Respondents rank profiles from most to least preferred. Provides more data per task but higher cognitive load."
  3. "MaxDiff (Best-Worst)" - "Respondents pick the best and worst from a set. Good for measuring attribute importance directly."

1d. Design Parameters

Question: "How many choice tasks and profiles per task?"
Header: "Design size"
Options:
  1. "Standard (8 tasks, 3 profiles) (Recommended)" - "Good balance of statistical power and respondent burden"
  2. "Compact (6 tasks, 2 profiles)" - "Fewer tasks for lower respondent burden, still adequate power"
  3. "Rich (12 tasks, 4 profiles)" - "More data per respondent, higher cognitive load"
  4. "Custom" - "Specify your own number of tasks and profiles per task"

If Custom, ask for specific numbers.

Then ask about the "none" option:

Question: "Should respondents have a 'None of these' option in each choice task?"
Header: "None option"
Options:
  1. "Yes" - "Include 'None of these' — useful when non-purchase is a realistic outcome"
  2. "No (Recommended)" - "Force a choice between the presented profiles"

1e. Respondent Segments

Question: "How would you like to define respondents?"
Header: "Respondents"
Options:
  1. "Define segments" - "Specify respondent segments with distinct traits (e.g., budget-conscious, tech enthusiast)"
  2. "Generic respondents" - "Use a default set of diverse respondents without specific segments"
  3. "Skip agents" - "Run without agent personas (default LLM behavior)"

If "Define segments", ask for segment descriptions:

Question: "Describe the respondent segments you want. For each segment, provide a name and key traits. For example:\n\nBudget-conscious: price-sensitive, value-oriented, income=$50k\nTech enthusiast: early adopter, feature-focused, income=$100k\nCasual user: convenience-oriented, low engagement"
Header: "Segments"

1f. Confirm Full Design

Present a complete summary:

Question: "Here is your conjoint study design:\n\n| Parameter | Value |\n|-----------|-------|\n| Method | Choice-Based Conjoint |\n| Attributes | <N> |\n| Total unique profiles | <N> |\n| Tasks per respondent | <N> |\n| Profiles per task | <N> |\n| Include 'None' option | Yes/No |\n| Respondent segments | <N> segments |\n| Design versions | 4 |\n| Total observations | tasks × versions × segments |\n\nReady to proceed?"
Header: "Confirm"
Options:
  1. "Proceed" - "Generate the experimental design and EDSL project files"
  2. "Modify" - "Go back and change some parameters"
  3. "Save design doc only" - "Save the design specification without generating code"

Phase 2: Generate Experimental Design

1. Write the design spec as design_spec.json in the project directory:

{
  "attributes": {
    "price": ["$9.99", "$14.99", "$19.99"],
    "content_library": ["Small", "Medium", "Large"],
    "ad_experience": ["No ads", "Limited ads", "Full ads"]
  },
  "method": "cbc",
  "tasks_per_version": 8,
  "profiles_per_task": 3,
  "n_versions": 4,
  "include_none": false,
  "min_attribute_diff": 2
   

2. Run the design generator:

python3 <helpers_path> generate-design <dir_name>/design_spec.json --output <dir_name>/conjoint_choice_sets.json

3. Write conjoint_design.md documenting the full design — research question, attributes table, method, parameters, and choice set summary.

Phase 3: Create EDSL Project Files

Generate a study directory following the create-study pattern. Before generating survey code, read the edsl-survey-reference skill for question type details:

Use Glob("**/edsl-survey-reference/SKILL.md") and read it for QuestionMultipleChoice details.

Project Structure

<dir_name>/
  conjoint_design.md          # Design document
  design_spec.json            # Machine-readable specification
  conjoint_choice_sets.json   # Generated choice sets
  study_survey.py             # Survey with choice task questions
  study_scenario_list.py      # Each design version as a Scenario
  study_agent_list.py         # Respondent segments
  study_model_list.py         # LLM models
  create_results.py           # Runner script
  analyze_conjoint.py         # Conjoint-specific analysis
  Makefile                    # Build targets

study_survey.py

Each choice task is a separate QuestionMultipleChoice. Profile descriptions are embedded in question text via Jinja2 scenario variables. Option labels are static ("Option A", "Option B", etc.).

from edsl import Survey, QuestionMultipleChoice

# === CHOICE TASKS ===

q_choice_1 = QuestionMultipleChoice(
    question_name="choice_task_1",
    question_text="""Consider the following options for a streaming service:

Option A: {{ task_1_opt_a }}
Option B: {{ task_1_opt_b }}
Option C: {{ task_1_opt_c }}

Which option would you choose?""",
    question_options=["Option A", "Option B", "Option C"]
)

q_choice_2 = QuestionMultipleChoice(
    question_name="choice_task_2",
    question_text="""Consider the following options for a streaming service:

Option A: {{ task_2_opt_a }}
Option B: {{ task_2_opt_b }}
Option C: {{ task_2_opt_c }}

Which option would you choose?""",
    question_options=["Option A", "Option B", "Option C"]
)

# ... repeat for all N tasks

# === SURVEY ===

survey = Survey([q_choice_1, q_choice_2, ...])

Key design decisions:

• Static option labels ("Option A", "Option B", "Option C") as question_options — NOT the profile descriptions
• Scenario templating for profile descriptions: each {{ task_N_opt_X }} is a formatted string describing one profile's attributes
• If include_none is True, add "None of these" to question_options
• Do NOT use set_full_memory_mode() — choice tasks are independent

study_scenario_list.py

Each design version becomes a separate Scenario. Scenario variables encode the formatted profile descriptions for every task and option.

from edsl import Scenario, ScenarioList

# Each scenario encodes one complete design version.
# Variables: task_{t}_opt_{a|b|c} = formatted profile description

scenario_list = ScenarioList([
    Scenario({
        "design_version": 1,
        "task_1_opt_a": "Price: $9.99, Content: Large, Ads: Full ads",
        "task_1_opt_b": "Price: $19.99, Content: Small, Ads: No ads",
        "task_1_opt_c": "Price: $14.99, Content: Medium, Ads: Limited ads",
        "task_2_opt_a": "...",
        # ... all tasks for version 1
    }),
    Scenario({
        "design_version": 2,
        # ... all tasks for version 2
    }),
    # ... more versions
])

To build the scenario list, read conjoint_choice_sets.json and format each profile as a multi-attribute description string. Also include individual attribute columns for analysis:

# For each task/option, also store individual attributes for analysis:
# task_{t}_opt_{x}_{attr_name} = level value

The scenario list construction should be done programmatically in the generated file:

import json
from edsl import Scenario, ScenarioList

with open("conjoint_choice_sets.json", "r") as f:
    design = json.load(f)

attributes = design["attributes"]
attr_names = list(attributes.keys())

scenarios = []
for version in design["versions"]:
    s = {"design_version": version["version"]}
    for t_idx, choice_set in enumerate(version["choice_sets"], 1):
        for p_idx, profile in enumerate(choice_set):
            opt_letter = chr(ord("a") + p_idx)
            # Formatted description for question text
            desc = ", ".join(f"{attr}: {profile[attr]}" for attr in attr_names)
            s[f"task_{t_idx}_opt_{opt_letter}"] = desc
            # Individual attributes for analysis
            for attr in attr_names:
                s[f"task_{t_idx}_opt_{opt_letter}_{attr}"] = profile[attr]
    scenarios.append(Scenario(s))

scenario_list = ScenarioList(scenarios)

study_agent_list.py

If segments were defined, create agents with traits. If generic respondents, create a diverse set. If skipped, export empty AgentList.

from edsl import Agent, AgentList

agent_list = AgentList([
    Agent(traits={"persona": "budget-conscious consumer", "segment": "budget", "income": "$50k"}),
    Agent(traits={"persona": "tech enthusiast", "segment": "tech", "income": "$100k"}),
    Agent(traits={"persona": "casual user", "segment": "casual", "income": "$75k"}),
])

study_model_list.py

from edsl import ModelList, Model

model_list = ModelList([Model("gpt-4o")])

create_results.py

from study_survey import survey
from study_scenario_list import scenario_list
from study_agent_list import agent_list
from study_model_list import model_list

job = survey.by(scenario_list).by(agent_list).by(model_list)
results = job.run()
results.to_json("results.json.gz")
results.to_csv("results.csv")

analyze_conjoint.py

A script that calls helpers.py analyze and then generates charts:

import subprocess
import sys

# Run conjoint analysis
subprocess.run([
    sys.executable, "<helpers_path>",
    "analyze", "results.csv", "design_spec.json",
    "--output-dir", "."
], check=True)

# Generate charts
subprocess.run([sys.executable, "generate_charts.py"], check=True)

print("Conjoint analysis complete. See:")
print("  conjoint_report.md  - Full analysis report")
print("  utilities.json      - Part-worth utilities")
print("  utilities.csv       - Utilities in CSV format")
print("  importance_chart.png - Attribute importance chart")
print("  utilities_chart.png  - Part-worth utilities chart")

Important: When generating analyze_conjoint.py, replace <helpers_path> with the actual discovered path to helpers.py.

Makefile

results.json.gz: create_results.py study_survey.py study_scenario_list.py study_agent_list.py study_model_list.py
	python create_results.py

analyze: results.csv analyze_conjoint.py design_spec.json
	python analyze_conjoint.py

results.csv: results.json.gz
	@echo "results.csv is produced alongside results.json.gz"

Phase 4: Run Study (Optional)

Ask the user:

Question: "The conjoint study files have been generated in <dir_name>/. Would you like to run the study now?"
Header: "Run study"
Options:
  1. "Run now" - "Execute the study (this will make API calls to LLM providers)"
  2. "Run later" - "I'll run it manually with 'make' or 'python create_results.py'"

If "Run now":

cd <dir_name> && python3 create_results.py

Phase 5: Analyze Results

After the study completes (or if results already exist):

1. Run the analysis:

cd <dir_name> && python3 analyze_conjoint.py

2. Present key findings from conjoint_report.md:

   • Attribute importance rankings
   • Top part-worth utilities
   • Segment differences (if applicable)

3. Suggest next steps:

Question: "The conjoint analysis is complete. What would you like to do next?"
Header: "Next steps"
Options:
  1. "Generate full report with /analyze-conjoint-results (Recommended)" - "Comprehensive HTML report with executive summary, methodology, charts, segment analysis, and limitations"
  2. "Run market simulation" - "Predict choice shares for specific product profiles"
  3. "Explore further with /analyze-results" - "Do additional generic analysis on the raw results"
  4. "Publish with /publish-study" - "Share the study on GitHub"

If "Run market simulation", ask for competing product profiles and run:

python3 <helpers_path> market-sim <dir_name>/utilities.json <dir_name>/profiles.json

Design Principles

1. Balanced designs: The randomized search optimizer ensures each attribute level appears approximately equally often across all choice tasks
2. Positional debiasing: Profile presentation order is shuffled across design versions to mitigate LLM positional bias
3. Multiple versions: 4 design versions by default increases effective sample size and reduces order effects
4. Minimum attribute differences: Profiles within a task differ on at least 2 attributes, forcing meaningful trade-offs
5. Static option labels: Using "Option A/B/C" rather than profile descriptions as answer choices ensures clean response parsing
6. Scenario-based variation: Each design version is a separate Scenario, leveraging EDSL's built-in parallelism

Output

The skill creates a project directory with all files needed to run and analyze a conjoint study:

Path	Description
conjoint_design.md	Human-readable design document
design_spec.json	Machine-readable design specification
conjoint_choice_sets.json	Generated balanced choice sets
study_survey.py	EDSL Survey with choice task questions
study_scenario_list.py	ScenarioList encoding design versions
study_agent_list.py	AgentList with respondent segments
study_model_list.py	ModelList specifying LLMs
create_results.py	Runner script
analyze_conjoint.py	Analysis script producing utilities and charts
Makefile	Build targets for running and analyzing