Generate a comprehensive, self-contained analysis report for a conjoint (choice-based) study with executive summary, methodology, part-worth utilities, segment analysis, and limitations
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Generate a comprehensive, self-contained analysis report for a conjoint (choice-based) study with executive summary, methodology, part-worth utilities, segment analysis, and limitations
If any attribute name contains "price" (case-insensitive), generate a dedicated price chart:
price_attr = Nonefor attr_name in attributes:
if"price"in attr_name.lower():
price_attr = attr_name
breakif price_attr:
price_utils = utilities[price_attr]
# Sort levels by numeric value extracted from the level stringimport re as re_mod
defextract_number(s):
m = re_mod.search(r'[\d.]+', s)
returnfloat(m.group()) if m else0
levels_sorted = sorted(price_utils.keys(), key=extract_number)
vals = [price_utils[l] for l in levels_sorted]
fig, ax = plt.subplots(figsize=(8, 5))
ax.plot(levels_sorted, vals, marker="o", linewidth=2, color="#4C78A8", markersize=8)
ax.fill_between(range(len(levels_sorted)), vals, alpha=0.1, color="#4C78A8")
ax.set_xlabel(price_attr.replace("_", " ").title())
ax.set_ylabel("Part-Worth Utility")
ax.set_title("Price Sensitivity")
ax.axhline(y=0, color="gray", linewidth=0.5)
ax.set_xticks(range(len(levels_sorted)))
ax.set_xticklabels(levels_sorted)
for i, (level, val) inenumerate(zip(levels_sorted, vals)):
ax.annotate(f"{val:+.3f}", (i, val), textcoords="offset points",
xytext=(0, 10), ha="center")
plt.tight_layout()
plt.savefig(os.path.join(output_dir, "price_sensitivity_chart.png"), dpi=150, bbox_inches="tight")
plt.close()
3e. Segment Importance Heatmap (if segment data exists)
if segment_data and"segment"in segment_data:
seg_importance = segment_data["segment"]
segments = sorted(seg_importance.keys())
all_attrs = attrs_sorted # use importance-sorted order
heatmap_data = []
for seg in segments:
row = []
seg_imp = seg_importance[seg]["importance"]
for attr in all_attrs:
row.append(seg_imp.get(attr, 0))
heatmap_data.append(row)
heatmap_array = np.array(heatmap_data)
fig, ax = plt.subplots(figsize=(max(8, len(all_attrs) * 1.5), max(5, len(segments) * 0.7)))
im = ax.imshow(heatmap_array, cmap="YlOrRd", aspect="auto")
ax.set_xticks(range(len(all_attrs)))
ax.set_xticklabels([a.replace("_", " ").title() for a in all_attrs], rotation=45, ha="right")
ax.set_yticks(range(len(segments)))
ax.set_yticklabels([s.replace("_", " ").title() for s in segments])
# Add text annotationsfor i inrange(len(segments)):
for j inrange(len(all_attrs)):
val = heatmap_array[i, j]
color = "white"if val > heatmap_array.max() * 0.7else"black"
ax.text(j, i, f"{val:.1f}%", ha="center", va="center", color=color, fontsize=9)
ax.set_title("Attribute Importance by Segment (%)")
plt.colorbar(im, label="Importance (%)")
plt.tight_layout()
plt.savefig(os.path.join(output_dir, "segment_heatmap.png"), dpi=150, bbox_inches="tight")
plt.close()
Phase 4: Build the Report
Build report.md with the following sections. Claude must write all prose sections — do not use placeholder text.
Report Structure
# Conjoint Analysis Report: [Research Question / Product Category]*Generated: YYYY-MM-DD HH:MM:SS*## Executive Summary
[Claude writes 1 paragraph stating the research question, the method used, the sample, and 3-5 quantified key findings. End with the optimal product configuration.]
**Optimal product configuration:** [List each attribute and its highest-utility level]
## Study Design### Attributes and Levels
| Attribute | Levels |
|-----------|--------|
| [attr1] | [level1, level2, ...] |
| ... | ... |
### Design Parameters
| Parameter | Value |
|-----------|-------|
| Method | Choice-Based Conjoint (CBC) |
| Tasks per respondent | [N] |
| Profiles per task | [N] |
| Design versions | [N] |
| Include "None" option | [Yes/No] |
| Total unique profiles | [N] |
| Total choice observations | [N] |
### Example Choice Tasks
[Show 1-2 REALIZED choice tasks from conjoint_choice_sets.json — not raw templates.
Format as the respondent would see them:]
> **Choice Task 1 (Version 1):**>
> | Attribute | Option A | Option B | Option C |> |-----------|----------|----------|----------|> | Price | $6.99 | $12.99 | $8.99 |> | Pack Size | 24 wings | 8 wings | 16 wings |> | ... | ... | ... | ... |>
> *Which option would you choose?*
[Show a second example from a different task or version.]
### Respondent Segments
[For EACH agent, show the segment label as a header and the FULL persona description. Never truncate.]
| Segment | Persona Description |
|---------|-------------------|
| [segment_label] | [Full persona text — do not truncate] |
| ... | ... |
**Total respondents:** [N segments] x [N design versions] = [N total response sets]
## Methodology
### Choice-Based Conjoint Analysis
Choice-Based Conjoint (CBC) is a stated-preference methodology where respondents evaluate sets of product profiles defined by multiple attributes and choose their preferred option. By observing which attribute combinations are chosen, we estimate the relative value (part-worth utility) each attribute level contributes to overall preference.
### Experimental Design
The experimental design was generated using a randomized search algorithm that optimizes for:
- **Level balance:** Each attribute level appears approximately equally often
- **Minimum attribute differences:** Profiles within each choice task differ on at least [N] attributes
- **Position debiasing:** Profile presentation order is shuffled across design versions
### Utility Estimation
Part-worth utilities are computed using a counting-based approach:
Where `choice_share` is the proportion of times a level was present in the chosen profile, and `expected_share` is 1/(number of levels). Utilities are zero-centered within each attribute.
**Attribute importance** is computed as the range of utilities within an attribute divided by the sum of all ranges:
### LLM Agent Simulation
Respondents in this study are simulated using large language model (LLM) agents, each assigned a distinct persona. Each agent independently evaluates the choice tasks based on its persona characteristics. This approach enables rapid, cost-effective preference estimation, though results should be validated against human respondent data for high-stakes decisions.
## Design Quality
### Position Bias
[Claude writes 1-2 sentences interpreting the position bias chart. Are options chosen roughly equally? Any concerning bias?]
### Design Balance
| Version | Balance Score |
|---------|--------------|
| [1] | [score] |
| ... | ... |
[Claude writes 1 sentence interpreting balance scores. Lower is better.]
## Overall Results
### Attribute Importance
| Attribute | Importance (%) |
|-----------|---------------|
| [attr] | [value] |
| ... | ... |
[Claude writes 2-3 sentences interpreting the importance rankings.]
### Part-Worth Utilities
[For EACH attribute, show a utility table:]
#### [Attribute Name] ([importance]%)
| Level | Utility |
|-------|---------|
| [level] | [+/-value] |
| ... | ... |
[Claude writes 1-2 sentences interpreting this attribute's utilities.]
### Optimal Product Configuration
Based on the highest part-worth utility for each attribute:
| Attribute | Optimal Level | Utility |
|-----------|--------------|---------|
| [attr] | [best level] | [value] |
| ... | ... | ... |
## Price Sensitivity
[ONLY include this section if a price attribute exists.]
[Claude writes 2-3 sentences about the price-utility relationship. Is it monotonically decreasing? Any interesting non-linearities?]
## Segment Analysis
### Importance Comparison Across Segments
[Claude writes 2-3 sentences comparing which segments weight which attributes most heavily.]
### Per-Segment Utilities
[For EACH segment, show:]
#### [Segment Label]: [Full Persona Description]
**Top priorities:** [List top 2-3 attributes by importance for this segment]
| Attribute | Importance (%) |
|-----------|---------------|
| ... | ... |
**Optimal product for this segment:**
| Attribute | Preferred Level | Utility |
|-----------|----------------|---------|
| ... | ... | ... |
[Claude writes 1-2 sentences about what distinguishes this segment.]
## Key Findings
[Claude writes 4-6 substantive paragraphs. Each paragraph should make a specific, quantified claim and explain its practical implications. Example topics:
1. Which attribute dominates preference and by how much
2. Price sensitivity patterns and willingness-to-pay implications
3. The most/least preferred levels within key attributes
4. Segment differences — which segments diverge most and on what
5. The optimal product configuration and how much utility it captures
6. Surprising findings or counter-intuitive patterns]
## Limitations
1. **LLM agent simulation:** Respondents are AI agents, not human consumers. Preferences may not perfectly reflect real-world purchase behavior. Results should be validated with human respondent studies before making high-stakes business decisions.
2. **Sample size:** With [N] design versions and [N] agent segments, the effective sample size is [N] response sets. Larger samples would increase confidence in utility estimates.
3. **Ecological validity:** Choice tasks present simplified attribute descriptions. Real purchase decisions involve additional factors (brand loyalty, shelf placement, availability) not captured here.
4. **Counting-based utilities:** The utility estimation uses a simple counting method rather than hierarchical Bayes or mixed logit. This may underestimate heterogeneity within segments.
5. **Attribute independence:** The model assumes attributes contribute independently to utility. Interaction effects (e.g., price × brand) are not estimated.
## Files Generated
| File | Description |
|------|-------------|
| [report.md](report.md) | This report in Markdown format |
| [report.html](report.html) | Styled HTML version of this report |
| [utilities.json](utilities.json) | Part-worth utilities and importance weights |
| [utilities.csv](utilities.csv) | Utilities in tabular format |
| [segment_analysis.json](segment_analysis.json) | Per-segment utility estimates |
| [importance_chart.png](importance_chart.png) | Attribute importance bar chart |
| [utilities_chart.png](utilities_chart.png) | Part-worth utilities by attribute |
| [position_bias_chart.png](position_bias_chart.png) | Position bias check |
| [price_sensitivity_chart.png](price_sensitivity_chart.png) | Price-utility curve (if applicable) |
| [segment_heatmap.png](segment_heatmap.png) | Cross-segment importance comparison |
Phase 5: Write Report and Convert to HTML
Write report.md to the output directory. Claude must fill in ALL prose sections with substantive content based on the actual data — no placeholders.
All prose sections are filled in (no placeholder text)
Persona descriptions are complete (not truncated)
Phase 6: Follow-Up
After generating the report, ask:
Question: "The conjoint analysis report has been generated at <output_dir>/report.html. What would you like to do next?"
Header: "Next steps"
Options:
1. "Generate PowerPoint" - "Create a slide deck summarizing the key findings"
2. "Run market simulation" - "Predict choice shares for specific product configurations"
3. "Publish with /publish-study" - "Share the study and report"
4. "Done" - "No further analysis needed"
If "Run market simulation", ask the user to define competing product profiles and run:
All chart generation must be done in a single Python script executed via Bash, or in sequential Bash calls. Do NOT attempt to run matplotlib interactively.
The helpers.py analyze command is the source of truth for utility computation. Do not reimplement the counting logic.
Always copy images into the output directory. Never use ../ paths in the report.
The report must be fully self-contained: a reader should understand the study without opening any other files.
Use segment trait values (not agent names) as display labels throughout.
Output
The skill creates an analysis_N/ directory inside the study directory containing:
