A Claude Code plugin that runs multi-agent user testing. Spawn AI persona agents — each with distinct demographics, behaviors, and a personal diary — to test your spec, web service, or API. Get a structured feedback report with prioritized recommendations.
claude plugin marketplace add cosmos-makers/user-test
claude plugin install user-test@user-test| Command | Description |
|---|---|
/user-test:setup [preset|custom] |
Configure persona group (preset or custom) |
/user-test:start <target> |
Run UT with configured personas → feedback report |
/user-test:enhance [report] |
Apply report recommendations to code/spec |
/user-test:run <target> [preset] |
One-shot: setup → start → enhance |
Automatically detected by target type:
- spec — Read and review documents, specs, or codebases
- web — Interact with live web services via browser (requires chrome MCP)
- api — Call API endpoints directly
| Preset | Description |
|---|---|
mz-2030 |
20-30s digital natives (3 personas) |
office-worker-40s |
40s office workers, varied IT proficiency (3 personas) |
senior |
50s+ seniors, accessibility focus (3 personas) |
diverse-mix |
Age/gender/proficiency balanced mix (4 personas) |
{
"id": "unique-id",
"label": "Display Name",
"demographics": { "age": 27, "gender": "male", "region": "서울", "income": "중상", "occupation": "개발자" },
"behavioral": { "domain_expertise": "high", "it_proficiency": "heavy_user", "competitor_experience": [], "current_workaround": "" },
"psychographic": { "tech_adoption": "early_adopter", "motivation": "", "usage_context": "" }
}.user-test/personas.json— Configured persona group.user-test/diaries/{id}.md— Persona diary (generated during setup).user-test/report-{timestamp}.md— Feedback report with prioritized recommendations
Research-backed enhancements to improve persona agent quality and realism.
Before evaluation, each agent explicitly enters the persona's perspective by answering: "What situation would I encounter this service in?", "What's my current state?", "What catches my eye in 3 seconds?"
This technique is shown to improve Theory-of-Mind capabilities in LLMs.
Agent prompts now include explicit instructions to base responses on defined behavioral traits and experiences, not demographic stereotypes. For example, a 52-year-old is not assumed to be bad with technology — instead, the agent references their specific it_proficiency: light_user and current_workaround: phone/in-person attributes.
After completing all evaluations, each agent performs a consistency check: re-reads their profile, verifies each response aligns with their defined attributes, corrects inconsistencies, and assigns a persona_consistency score (1-5). Low scores are flagged in the report.
During setup, each persona gets an auto-generated "yesterday's diary" — a 300-500 character first-person journal entry reflecting their daily life, digital service usage patterns, and frustrations. This provides vivid experiential context that makes test responses more grounded and realistic.
Diaries are stored in .user-test/diaries/{id}.md and automatically loaded as context during testing.
These improvements are grounded in the following academic research on LLM role-playing, persona simulation, and character consistency:
| # | Paper | Year | Venue | Applied In |
|---|---|---|---|---|
| 1 | Shanahan et al. "Role Play with Large Language Models" | 2023 | Nature | Theoretical foundation |
| 2 | Shao et al. "Character-LLM: A Trainable Agent for Role-Playing" | 2023 | EMNLP | Diary (experience replay) |
| 3 | Park et al. "Generative Agents: Interactive Simulacra of Human Behavior" | 2023 | UIST | Memory/reflection architecture |
| 4 | Gupta et al. "Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs" | 2024 | ICLR | Anti-stereotype guardrail |
| 5 | Zheng et al. "Quantifying Global Faithfulness in Persona-Driven Role-Playing" | 2024 | NeurIPS | Consistency scoring (APC) |
| 6 | Cai et al. "Persona-Aware Contrastive Learning for Persona Consistency" | 2025 | ACL | Chain-of-Persona |
| 7 | "UXAgent: An LLM Agent-Based Usability Testing Framework" | 2025 | CHI | Dual-system cognitive model |
| 8 | "Think Twice: Perspective-Taking Improves LLMs' ToM Capabilities" | 2023 | arXiv | Perspective-taking prompting |
| 9 | Kim et al. "Persona is a Double-Edged Sword" | 2024 | arXiv | Ensemble validation |
| 10 | "Human Simulacra: A Step toward Personification of LLMs" | 2024 | arXiv | Diary (life story generation) |
Full survey covers 23 papers on LLM role-playing, persona simulation, and character consistency.
MIT
