AudienceLens Harness
jedilpf/audiencelens-harness
Simule painéis de audiência sintéticos com avaliação estruturada em 8 dimensões
Ver no GitHubPack de skills para multi-agentes que simula painéis sintéticos de audiência para avaliar games, produtos e ideias antes de testes reais. Inclui 4 skills de review com sistema de qualidade de 8 dimensões, personas comportamentais e red-teaming adversarial.
Instalação
npx convez-ai jedilpf/audiencelens-harness1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Validar conceitos de games e produtos sem usuários reais
- Identificar riscos e conflitos de segmentação em landing pages e SaaS
- Red-teaming adversarial para cenários de falha e colapso de confiança
- Gerar personas reutilizáveis baseadas em comportamento para painéis customizados
Veredito

Ferramenta bem estruturada para síntese de feedback de audiência com rigor metodológico (8 dimensões de qualidade, personas comportamentais, red-teaming). Útil para validação pré-lançamento, mas requer curva de aprendizado e interpretação cuidadosa dos limites de feedback sintético vs. real.
README
A tradução para o português deste README ainda está na fila.
AudienceLens Harness
Open-Source Multi-Agent Synthetic Audience Review Skill Pack with Quality Harness
开源多智能体合成观众评审技能包 · 含8维质量评估体系
Overview
AudienceLens Harness is a self-contained, zero-dependency skill pack that enables Claude Code / Trae AI / Cursor-style agents to simulate diverse audience panels for reviewing games, products, landing pages, DevLogs, startup ideas, and creative projects — before committing resources to real user testing.
Unlike simply asking an LLM for feedback, AudienceLens runs multiple behavior-driven personas in isolation, then applies structured cross-examination on top-3 controversies, tracks stance revisions with full provenance, and validates output quality through an 8-dimension Harness.
Feature Categories
1. Skill System — Four Review Modes
| Skill | Command | Use Case |
|---|---|---|
| synthetic-playtest | /synthetic-playtest <input> [--focus] [--panel] [--count] [--harness] | Game-focused synthetic player panel — Steam pages, demos, concepts, monetization |
| audience-panel | /audience-panel <input> [--focus] [--panel] [--count] [--harness] | General product / app / landing-page / SaaS audience review |
| red-team-audience | /red-team-audience <input> [--intensity] [--mode] [--harness] | Adversarial stress test — failure modes, worst-case scenarios, trust collapse |
| persona-generate | /persona-generate <input> [--count] [--market] | Generate reusable persona panels as schema-compliant YAML |
2. Harness Quality Gate — 8 Evaluation Dimensions
The Harness is what differentiates AudienceLens from a prompt pack. It provides a repeatable, rubric-based quality check for every generated report.
| # | Dimension | What It Checks | Score Range |
|---|---|---|---|
| 1 | Persona Differentiation | Do personas produce distinct perspectives, or are they interchangeable? | 0–3 |
| 2 | Input Grounding | Is feedback anchored in actual input materials, or is it generic advice? | 0–3 |
| 3 | Actionability | Does the report include concrete P0/P1/P2 fixes, or only vague suggestions? | 0–3 |
| 4 | Segment Conflict | Are genuine disagreements between audience segments surfaced? | 0–3 |
| 5 | Risk Detection | Are critical risks and blockers identified and prioritized? | 0–3 |
| 6 | Ethical Persona Use | Are demographics/MBTI used as weak modifiers only — never as deterministic stereotypes? | 0–3 |
| 7 | Report Structure | Does the report include all required sections per the schema? | 0–3 |
| 8 | Calibration Honesty | Does the report frame findings as hypotheses, not predictions? | 0–3 |
Critical Failures (automatic fail, regardless of total score):
- No AI-generated disclaimer present
- No P0/P1/P2 prioritized actions
- Personas produce identical feedback
- Feedback is not grounded in input
- Gender/MBTI used as deterministic stereotype
- Claims real-world success prediction
3. Persona System — Behavior-Driven Design
Personas are defined by behavioral vectors, not demographic labels. Research (SCOPE study) shows demographics explain only 1.5% of behavioral variance.
Each persona includes:
- Behavior profile — patience, domain knowledge, novelty seeking, risk tolerance, etc.
- Decision rules — explicit criteria for positive/negative evaluation
- Quit triggers — conditions under which the persona would disengage
- Continue triggers — conditions that sustain engagement
- Must-comment-on — mandatory evaluation topics per persona role
4. Ethics Framework — Built-In Safeguards
| Principle | Implementation |
|---|---|
| Identity-Behavior Decoupling | Gender, age, region are weak style modifiers only — never causal predictors |
| MBTI De-authorization | MBTI modifies expression style only — never core judgment logic |
| Evidence Tiering | [direct] / [inferred] / [gap] — inferred evidence capped at 40% |
| No Fabrication | Personas must not invent missing features, plot, multiplayer, or monetization details |
| Mandatory Disclaimer | Every report includes "AI-generated synthetic review — not real user research" |
5. Panel Presets — Pre-Configured Audience Panels
| Preset | File | Composition |
|---|---|---|
| game-mixed | persona-panels/game-mixed.yaml | Balanced mix of core, casual, and critical game personas |
| game-steam-page | persona-panels/game-steam-page.yaml | Steam browsing behavior specialists |
| red-team-default | persona-panels/red-team-default.yaml | Adversarial attackers for stress testing |
6. Report Templates — Structured Output
| Template | Skill | Key Sections |
|---|---|---|
| synthetic-playtest-report.md | synthetic-playtest | Executive conclusion, player panel, consensus findings, risks & opportunities, priority fixes |
| audience-panel-report.md | audience-panel | Segment summary, consensus, disagreements, actionable recommendations |
| red-team-audience-report.md | red-team-audience | Must-fix issues, can-ignore items, counter-measures, worst-case scenarios |
Quick Start
Install as Claude Code Skill
# Copy skill files to your project's Claude config
cp -r .claude/skills/* /path/to/your-project/.claude/skills/
# Or install globally (all projects)
cp -r .claude/skills/* ~/.claude/skills/
# Windows (PowerShell)
# Copy-Item -Recurse .claude/skills/* -Destination $env:USERPROFILE\.claude\skills\
Run a Review
# Game review with mixed panel
/synthetic-playtest examples/game-demo-review-input.md --focus onboarding --panel mixed
# Product / landing page review
/audience-panel examples/landing-page-review-input.md --focus positioning
# Red-team adversarial review
/red-team-audience examples/game-demo-review-input.md --intensity high
# Generate custom personas
/persona-generate --count 6 --market game
Harness Evaluation Workflow
Run skill on fixture → Generate report → Evaluate with rubric → Identify weak dimensions → Revise skill/persona/template → Re-run
Repository Structure
audiencelens-harness/
├── .claude/
│ ├── skills/ # 4 Claude Code Skills
│ │ ├── synthetic-playtest/ # Game synthetic playtest
│ │ ├── audience-panel/ # General product review
│ │ ├── red-team-audience/ # Red-team attack review
│ │ └── persona-generate/ # Persona YAML generator
│ └── docs/
│ ├── persona-panels/ # Panel presets (YAML)
│ │ ├── game-mixed.yaml
│ │ ├── game-steam-page.yaml
│ │ └── red-team-default.yaml
│ └── templates/ # Report templates (Markdown)
│ ├── synthetic-playtest-report.md
│ ├── audience-panel-report.md
│ └── red-team-audience-report.md
├── harness/ # Quality Harness
│ ├── evaluators/ # 5 evaluation criteria (Markdown)
│ │ ├── actionability.md
│ │ ├── input-grounding.md
│ │ ├── persona-differentiation.md
│ │ ├── stereotype-check.md
│ │ └── structure-check.md
│ ├── rubrics/ # Scoring rubric
│ │ └── audience-review-rubric.md
│ ├── expected/ # Required report sections
│ │ └── synthetic-playtest-required-sections.yaml
│ └── fixtures/ # Test fixtures
│ ├── game-concept-cyber-scavenger.md
│ └── weak-game-concept.md
├── personas/ # Persona definitions
│ ├── game/ # Game personas (YAML)
│ │ ├── core-indie-player.yaml
│ │ ├── steam-browser.yaml
│ │ └── visual-first-player.yaml
│ ├── schemas/ # Persona schema
│ │ └── persona.schema.yaml
│ └── ethics.md # Ethics guidelines
├── docs/ # Documentation
│ ├── methodology.md
│ ├── hosted-version.md
│ └── roadmap.md
├── examples/ # Example inputs
│ ├── game-demo-review-input.md
│ └── landing-page-review-input.md
├── README.md
├── LICENSE # MIT
└── .gitignore
Roadmap
| Version | Focus | Status |
|---|---|---|
| v0.1 | Skill Pack (Claude Code skills + Persona YAML + Templates) | Done |
| v0.2 | Harness Edition (Rubrics + Evaluators + Schema + Ethics + Self-check) | Current |
| v0.3 | CLI (audiencelens run, audiencelens harness evaluate) | Planned |
| v0.4 | MCP Server (5 tools for programmatic access) | Planned |
| v1.0 | Hosted Integration (Shuoji platform + visual reports) | Planned |
Hosted Version
| Open-source Edition | Shuoji Hosted Edition | |
|---|---|---|
| Setup | Local setup required | One-click |
| Model | Bring your own key | Managed |
| Output | Markdown / YAML | Visual interactive reports |
| History | Manual | Report history + version comparison |
| Sharing | Manual | Private/public sharing |
| Personas | Edit YAML manually | Advanced persona panel builder |
Platform feature name: 朔纪 AI 玩家评审团
Disclaimer
Synthetic audience feedback is not a substitute for real user research. Use it for:
- Blind-spot detection
- Hypothesis generation
- Pre-screening before real testing
- Critique preparation
- Next-test planning
License
MIT License — see LICENSE
Made by Shuoji Team
概述
AudienceLens Harness 是一个自包含、零依赖的技能包,让 Claude Code / Trae AI / Cursor 类 AI 代理能够模拟多元观众面板,在投入真人测试资源之前对游戏、产品、着陆页、DevLog、创业方案和创意项目进行评审。
与直接问 LLM 不同,AudienceLens 先隔离运行多个行为驱动人格,再对 top-3 争议点执行结构化交叉质询,追踪立场修正的完整溯源链,最后通过8维 Harness 验证输出质量。
功能分类
1. 技能系统 — 四种评审模式
| 技能 | 命令 | 适用场景 |
|---|---|---|
| synthetic-playtest | /synthetic-playtest <input> [--focus] [--panel] [--count] [--harness] | 游戏专项合成玩家面板 — Steam页、Demo、概念稿、商业化 |
| audience-panel | /audience-panel <input> [--focus] [--panel] [--count] [--harness] | 通用产品 / App / 着陆页 / SaaS 观众评审 |
| red-team-audience | /red-team-audience <input> [--intensity] [--mode] [--harness] | 对抗性压力测试 — 失败模式、最坏场景、信任崩塌 |
| persona-generate | /persona-generate <input> [--count] [--market] | 生成可复用的、符合 Schema 的人格面板 YAML |
2. Harness 质量门控 — 8 个评估维度
Harness 是 AudienceLens 区别于普通 Prompt 包的核心——它提供可重复的、基于量表的质量检查。
| # | 维度 | 检查内容 | 评分范围 |
|---|---|---|---|
| 1 | 人格差异化 | 人格是否产出不同视角,还是可以互换? | 0–3 |
| 2 | 输入依据 | 反馈是否锚定在实际输入材料上,还是泛泛而谈? | 0–3 |
| 3 | 可操作性 | 报告是否包含具体的 P0/P1/P2 修复项? | 0–3 |
| 4 | 群体冲突 | 观众群体之间的真实分歧是否被暴露? | 0–3 |
| 5 | 风险检测 | 关键风险和阻塞项是否被识别和优先排序? | 0–3 |
| 6 | 伦理人格使用 | 人口统计/MBTI 是否仅作为弱修饰符——而非决定性刻板印象? | 0–3 |
| 7 | 报告结构 | 报告是否包含 Schema 要求的所有段落? | 0–3 |
| 8 | 校准诚实度 | 报告是否将发现定位为假设而非预测? | 0–3 |
关键失败条件(无论总分如何,触发即判定不通过):
- 缺少 AI 生成免责声明
- 缺少 P0/P1/P2 优先级修复项
- 人格产出完全相同的反馈
- 反馈未基于输入材料
- 性别/MBTI 被用作决定性刻板印象
- 声称能预测真实世界结果
3. 人格系统 — 行为驱动设计
人格由行为向量定义,而非人口统计标签。研究(SCOPE)表明人口统计仅解释 1.5% 的行为方差。
每个人格包含:
- 行为画像 — 耐心、领域知识、创新偏好、风险容忍等
- 决策规则 — 正面/负面评价的显式标准
- 退出触发器 — 人格会放弃的条件
- 继续触发器 — 维持参与的条件
- 必评项 — 每个人格角色必须评价的主题
4. 伦理框架 — 内置安全机制
| 原则 | 实现 |
|---|---|
| 身份-行为解耦 | 性别、年龄、地域仅作为弱风格修饰符——绝不作为因果预测因子 |
| MBTI 去权 | MBTI 仅修饰表达风格——不推导核心判断逻辑 |
| 证据分级 | [direct] / [inferred] / [gap] — 推断证据上限 40% |
| 禁止脑补 | 人格不得编造缺失的功能、剧情、联机或商业化细节 |
| 强制免责 | 每份报告包含"AI 生成的合成评审——非真实用户研究" |
5. 面板预设 — 预配置观众面板
| 预设 | 文件 | 组成 |
|---|---|---|
| game-mixed | persona-panels/game-mixed.yaml | 核心、休闲、批判型游戏人格的均衡组合 |
| game-steam-page | persona-panels/game-steam-page.yaml | Steam 浏览行为专家 |
| red-team-default | persona-panels/red-team-default.yaml | 对抗性攻击者,用于压力测试 |
6. 报告模板 — 结构化输出
| 模板 | 技能 | 关键段落 |
|---|---|---|
| synthetic-playtest-report.md | synthetic-playtest | 执行结论、玩家面板、共识发现、风险与机会、优先修复项 |
| audience-panel-report.md | audience-panel | 群体摘要、共识、分歧、可操作建议 |
| red-team-audience-report.md | red-team-audience | 必修项、可忽略项、对策、最坏场景 |
快速开始
安装为 Claude Code Skill
# 复制到你的项目目录
cp -r .claude/skills/* /path/to/your-project/.claude/skills/
# 或全局安装(所有项目可用)
cp -r .claude/skills/* ~/.claude/skills/
# Windows (PowerShell)
# Copy-Item -Recurse .claude/skills/* -Destination $env:USERPROFILE\.claude\skills\
运行评审
# 游戏评审(混合面板)
/synthetic-playtest examples/game-demo-review-input.md --focus onboarding --panel mixed
# 产品 / 着陆页评审
/audience-panel examples/landing-page-review-input.md --focus positioning
# 红队对抗评审
/red-team-audience examples/game-demo-review-input.md --intensity high
# 生成自定义人格
/persona-generate --count 6 --market game
Harness 评估流程
在测试用例上运行技能 → 生成报告 → 用量表评估 → 识别弱维度 → 修订技能/人格/模板 → 重新运行
仓库结构
audiencelens-harness/
├── .claude/
│ ├── skills/ # 4 个 Claude Code 技能
│ │ ├── synthetic-playtest/ # 游戏合成评审
│ │ ├── audience-panel/ # 通用产品评审
│ │ ├── red-team-audience/ # 红队对抗评审
│ │ └── persona-generate/ # 人格 YAML 生成器
│ └── docs/
│ ├── persona-panels/ # 面板预设 (YAML)
│ │ ├── game-mixed.yaml
│ │ ├── game-steam-page.yaml
│ │ └── red-team-default.yaml
│ └── templates/ # 报告模板 (Markdown)
│ ├── synthetic-playtest-report.md
│ ├── audience-panel-report.md
│ └── red-team-audience-report.md
├── harness/ # 质量 Harness
│ ├── evaluators/ # 5 个评估标准 (Markdown)
│ │ ├── actionability.md
│ │ ├── input-grounding.md
│ │ ├── persona-differentiation.md
│ │ ├── stereotype-check.md
│ │ └── structure-check.md
│ ├── rubrics/ # 评分量表
│ │ └── audience-review-rubric.md
│ ├── expected/ # 报告必需段落
│ │ └── synthetic-playtest-required-sections.yaml
│ └── fixtures/ # 测试用例
│ ├── game-concept-cyber-scavenger.md
│ └── weak-game-concept.md
├── personas/ # 人格定义
│ ├── game/ # 游戏人格 (YAML)
│ │ ├── core-indie-player.yaml
│ │ ├── steam-browser.yaml
│ │ └── visual-first-player.yaml
│ ├── schemas/ # 人格 Schema
│ │ └── persona.schema.yaml
│ └── ethics.md # 伦理指南
├── docs/ # 文档
│ ├── methodology.md
│ ├── hosted-version.md
│ └── roadmap.md
├── examples/ # 示例输入
│ ├── game-demo-review-input.md
│ └── landing-page-review-input.md
├── README.md
├── LICENSE # MIT
└── .gitignore
路线图
| 版本 | 重点 | 状态 |
|---|---|---|
| v0.1 | 技能包(Claude Code 技能 + 人格 YAML + 模板) | 已完成 |
| v0.2 | Harness 版(量表 + 评估器 + Schema + 伦理 + 自检) | 当前 |
| v0.3 | CLI(audiencelens run、audiencelens harness evaluate) | 计划中 |
| v0.4 | MCP Server(5 个工具,支持程序化访问) | 计划中 |
| v1.0 | 托管集成(朔纪平台 + 可视化报告) | 计划中 |
托管版本
| 开源版 | 朔纪托管版 | |
|---|---|---|
| 安装 | 需本地配置 | 一键启动 |
| 模型 | 自备 API Key | 托管 |
| 输出 | Markdown / YAML | 可视化交互报告 |
| 历史记录 | 手动管理 | 报告历史 + 版本对比 |
| 分享 | 手动 | 私密/公开分享 |
| 人格管理 | 手动编辑 YAML | 高级人格面板构建器 |
平台功能名:朔纪 AI 玩家评审团
免责声明
合成观众反馈不能替代真实用户研究。 适用于:
- 盲点检测
- 假设生成
- 真实测试前的预筛选
- 评审准备
- 下一步测试规划
许可证
MIT License — 详见 LICENSE