Back to Details

t3-hardware-scoring

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

MantaBase T3 Hardware Audit System

MantaBase T3 Hardware Audit System

Objectives

目标

  • This Skill: Automatically crawls product information from hardware product links, applies Brand Blinding for debranding, has three independent Auditors score using authoritative reference checklists, performs Peer Review for mutual critique, and produces objective Tool/Toy/Trash classification
  • Capabilities: Web scraping, Brand Blinding filtering, Triple-Auditor specialized scoring, Peer Review, Final Judge synthesis
  • Triggers: User provides hardware product links requesting T3 audit or classification
  • Core principles: Transparency first, objective evaluation, parallel audits, expert review, continuous improvement
  • 本Skill：自动从硬件产品链接爬取产品信息，应用Brand Blinding去除品牌标识，由三位独立审计员使用权威参考清单评分，开展同行评审互评，最终生成客观的工具/玩具/垃圾分类
  • 功能：网页爬取、Brand Blinding过滤、三重审计员专项评分、同行评审、最终裁决合成
  • 触发条件：用户提供硬件产品链接并请求T3审计或分类
  • 核心原则：透明优先、客观评估、并行审计、专家评审、持续改进

Prerequisites

前置条件

  • No extra setup; the system handles web content extraction
  • 无需额外设置；系统会处理网页内容提取

Procedure

流程

Standard Flow

标准流程

1. Get User Input

1. 获取用户输入

  • Ask user for hardware product links (e-commerce, official sites, review sites, etc.)
  • Optional: Ask about specific dimensions or audit focus
  • 向用户索要硬件产品链接（电商平台、官网、评测网站等）
  • 可选：询问特定评估维度或审计重点

2. Get Product Information (Agent Method Recommended)

2. 获取产品信息（推荐使用Agent方法）

Required read references/mandatory-page-list.md. Execute the fixed crawl list by site type (Shopify / Kickstarter) for consistent data input.
Method A: Agent Direct Fetch (Recommended)
  • Read references/web-fetch-guide.md for full extraction strategy
  • Identify site type, then fetch per mandatory-page-list, write to 
    01-level0-source-urls.md
    , 
    02-level0-extracts.md
  • Use web_fetch tool to access product pages
  • Advantages:
    • ✅ High data completeness (smart extraction)
    • ✅ Adapts to different site structures
    • ✅ Handles complex pages (dynamic content, tabs)
    • ✅ Can visit multiple related pages (product, specs, reviews)
    • ✅ Agent can identify and fill gaps
  • Execution notes: 
    ⚠️ Important: A single page may be incomplete. Agent should:

1. Multi-page visits:
   - Product page → basic info, price
   - Specs page → technical params
   - Reviews page → user feedback
   - Review/editorial pages → professional evaluation

2. Price extraction:
   - Check schema.org structured data
   - Try cart/checkout pages
   - Search "product name + price"
   - If missing: state clearly in report

3. Dynamic content:
   - Use paginated fetch (offset params)
   - Try different URL variants
   - Check mobile pages

4. Data completeness:
   - Basic info: must be complete
   - Specs: at least 50%
   - User feedback: at least one source
Method B: Python Script (Bulk/Offline)
  • Use when: batch products, offline analysis
  • Call script to extract product content:
    bash
    python scripts/crawl_product_info.py --url <product_url> --pretty
  • Params: 
    --url
    (required), 
    --pretty
    (optional), 
    --output
    (optional)
  • Note: Script may miss data; prefer Agent method
  • Script returns: product name, description, features, price, target users, objective data, etc.
必读文档 references/mandatory-page-list.md。根据站点类型（Shopify / Kickstarter）执行固定爬取列表，确保数据输入一致性。
方法A：Agent直接获取（推荐）
  • 阅读 references/web-fetch-guide.md 了解完整提取策略
  • 识别站点类型，然后按照mandatory-page-list爬取内容，写入
    01-level0-source-urls.md
    、
    02-level0-extracts.md
  • 使用web_fetch工具访问产品页面
  • 优势：
    • ✅ 数据完整性高（智能提取）
    • ✅ 适配不同站点结构
    • ✅ 处理复杂页面（动态内容、标签页）
    • ✅ 可访问多个相关页面（产品页、参数页、评测页）
    • ✅ Agent可识别并填补数据缺口
  • 执行注意事项： 
    ⚠️ Important: A single page may be incomplete. Agent should:

1. Multi-page visits:
   - Product page → basic info, price
   - Specs page → technical params
   - Reviews page → user feedback
   - Review/editorial pages → professional evaluation

2. Price extraction:
   - Check schema.org structured data
   - Try cart/checkout pages
   - Search "product name + price"
   - If missing: state clearly in report

3. Dynamic content:
   - Use paginated fetch (offset params)
   - Try different URL variants
   - Check mobile pages

4. Data completeness:
   - Basic info: must be complete
   - Specs: at least 50%
   - User feedback: at least one source
方法B：Python脚本（批量/离线）
  • 适用场景：批量产品分析、离线分析
  • 调用脚本提取产品内容：
    bash
    python scripts/crawl_product_info.py --url <product_url> --pretty
  • 参数：
    --url
    （必填）、
    --pretty
    （可选）、
    --output
    （可选）
  • 注意：脚本可能遗漏数据，优先使用Agent方法
  • 脚本返回内容：产品名称、描述、特性、价格、目标用户、客观数据等

3. Brand Blinding (Agent)

3. Brand Blinding（由Agent执行）

  • Read references/defluff-guide.md for Brand Blinding rules
  • Agent performs debranding:
    • Remove brand names and trademark references
    • Remove marketing language and hype
    • Identify and remove emotional adjectives
    • Keep functional and objective information
    • Preserve factual accuracy
  • Output: Brand-Blinded objective product info
  • 阅读 references/defluff-guide.md 了解Brand Blinding规则
  • Agent执行去品牌化操作：
    • 移除品牌名称和商标引用
    • 移除营销话术和夸大宣传
    • 识别并移除情绪化形容词
    • 保留功能性和客观信息
    • 确保事实准确性
  • 输出：经过Brand Blinding处理的客观产品信息

4. Triple Auditor Specialized Scoring (Agent, Parallel)

4. 三重审计员专项评分（Agent并行执行）

  • Each Auditor evaluates independently and in parallel
  • Each Auditor reads only their own guide, not others’ guides or reports
🟢 Tool Auditor
  • Read references/tool-auditor.md
  • Follow references/auditor-templates.md strictly
  • Focus: Pain-point solving, practicality, reliability
  • Reference: Tony Fadell "Build" checklist
  • Dimensions:
    • Pain point identification (30 pts)
    • Attention to detail and consistency (25 pts)
    • Simplicity and efficiency (25 pts)
    • Engineering reliability (20 pts)
  • Output: Scoring report (100 pts total) with item scores, reasons, evidence
  • Litmus Test: If it broke tomorrow, would the user’s workflow stall?
🟡 Toy Auditor
  • Read references/toy-auditor.md
  • Follow references/auditor-templates.md strictly
  • Focus: Emotional value, enjoyment, aesthetic design
  • Reference: Don Norman "The Design of Everyday Things" checklist
  • Dimensions:
    • Sensory pleasure (30 pts)
    • Surprise and discovery (25 pts)
    • Emotional connection (25 pts)
    • Explorability (20 pts)
  • Output: Scoring report (100 pts total) with item scores, reasons, evidence
  • Litmus Test: Would the user put it on display or keep it as a collectible?
🔴 Trash Auditor
  • Read references/trash-auditor.md
  • Follow references/auditor-templates.md strictly
  • Required references/trash-red-flags.md high-sensitivity triggers
  • Focus: Logical flaws, marketing deception, design violations
  • Reference: Dieter Rams Ten Principles violation checklist
  • Dimensions:
    • Principle violations (30 pts)
    • Problem creation (25 pts)
    • Value deficit (25 pts)
    • Replaceability (20 pts)
  • Output: Scoring report (0–100 pts, higher = more Trash) with item scores, reasons, evidence
  • Litmus Test: If this product disappeared tomorrow, would the world be better or worse?
Key Principles:
  • Information isolation: Auditors only see Brand-Blinded info; never raw product info
  • Checklist-based: Must use the reference checklists
  • Evidence-driven: Every score backed by evidence
  • Independent: Three Auditors evaluate in parallel, no cross-talk
  • Objective: Report findings, not defend a category
  • Honest: Record evidence even when it supports another category
  • 每位审计员独立并行评估
  • 每位审计员仅阅读自身对应的指南，不得查看其他审计员的指南或报告
🟢 工具类审计员
  • 阅读 references/tool-auditor.md
  • 严格遵循 references/auditor-templates.md
  • 评估重点：痛点解决能力、实用性、可靠性
  • 参考依据：Tony Fadell《Build》清单
  • 评估维度：
    • 痛点识别（30分）
    • 细节关注与一致性（25分）
    • 简洁性与效率（25分）
    • 工程可靠性（20分）
  • 输出：评分报告（满分100分），包含分项得分、理由及证据
  • 测试标准：如果明天这个产品坏了，用户的工作流会停滞吗？
🟡 玩具类审计员
  • 阅读 references/toy-auditor.md
  • 严格遵循 references/auditor-templates.md
  • 评估重点：情感价值、趣味性、美学设计
  • 参考依据：Don Norman《设计心理学》清单
  • 评估维度：
    • 感官愉悦度（30分）
    • 惊喜与探索性（25分）
    • 情感联结（25分）
    • 可探索性（20分）
  • 输出：评分报告（满分100分），包含分项得分、理由及证据
  • 测试标准：用户会把它展示出来或作为收藏品保留吗？
🔴 垃圾类审计员
  • 阅读 references/trash-auditor.md
  • 严格遵循 references/auditor-templates.md
  • 必须参考 references/trash-red-flags.md 中的高敏感度触发项
  • 评估重点：逻辑缺陷、营销欺诈、设计违规
  • 参考依据：Dieter Rams十大设计原则违规清单
  • 评估维度：
    • 原则违规（30分）
    • 制造问题（25分）
    • 价值缺失（25分）
    • 可替代性（20分）
  • 输出：评分报告（0-100分，分数越高越接近垃圾类），包含分项得分、理由及证据
  • 测试标准：如果这个产品明天消失，世界会变得更好还是更糟？
核心原则:
  • 信息隔离：审计员仅能查看经过Brand Blinding处理的信息；不得查看原始产品信息
  • 基于清单：必须使用参考清单
  • 证据驱动：每一项评分都要有证据支持
  • 独立性：三位审计员并行评估，不得交叉沟通
  • 客观性：仅报告发现，无需为某一分类辩护
  • 诚实性：即使证据支持其他分类，也要如实记录

5. Peer Review (Agent, Parallel)

5. 同行评审（Agent并行执行）

  • Read references/peer-review-guide.md
  • Each Auditor reviews the other two reports:
Cross-review:
  • Tool reviews: Toy + Trash
  • Toy reviews: Tool + Trash
  • Trash reviews: Tool + Toy
Review focus:
  • Are scores reasonable?
  • Is evidence sufficient?
  • Any cross-category evidence missed?
  • Bias or omissions?
Output:
  • Review comments (with suggestions and reasons)
  • Score adjustment suggestions
  • Cross-category evidence
  • 阅读 references/peer-review-guide.md
  • 每位审计员评审另外两位的报告：
交叉评审规则:
  • 工具类审计员评审：玩具类 + 垃圾类报告
  • 玩具类审计员评审：工具类 + 垃圾类报告
  • 垃圾类审计员评审：工具类 + 玩具类报告
评审重点:
  • 评分是否合理？
  • 证据是否充分？
  • 是否遗漏跨分类证据？
  • 是否存在偏见或遗漏？
输出:
  • 评审意见（含建议及理由）
  • 评分调整建议
  • 跨分类证据

6. Auditor Report Optimization (Agent)

6. 审计员报告优化（Agent执行）

  • Each Auditor refines their report using review feedback:
    • Assess validity of feedback
    • Accept valid suggestions and update report
    • For invalid suggestions, give rebuttal
    • Log all changes and reasons
Output:
  • Original report
  • Optimized report
  • Change log (accepted/rejected comments and reasons)
  • 每位审计员根据评审反馈优化自身报告：
    • 评估反馈的有效性
    • 接受有效建议并更新报告
    • 对无效建议给出反驳理由
    • 记录所有变更及原因
输出:
  • 原始报告
  • 优化后报告
  • 变更日志（接受/拒绝的意见及理由）

7. Final Judge Synthesis (Agent)

7. 最终裁决合成（Agent执行）

  • Read references/t3-classification.md
  • Read all three optimized Auditor reports
  • Synthesize conclusions and evidence
  • Apply T3 rules:
    • Composite = max(Tool, Toy) - Trash
    • Determine primary category (≥2 conditions met)
    • Determine secondary category (if any)
    • Apply Litmus Test consistency check
  • Output: Final classification, confidence, reasoning, improvement suggestions
  • 阅读 references/t3-classification.md
  • 阅读三份优化后的审计员报告
  • 综合结论与证据
  • 应用T3规则：
    • 综合得分 = max(工具类得分, 玩具类得分) - 垃圾类得分
    • 确定主分类（满足≥2个条件）
    • 确定次分类（如有）
    • 执行测试标准一致性检查
  • 输出：最终分类结果、置信度、推理过程、改进建议

8. Generate Audit Report

8. 生成审计报告

Output file: Always 
99-audit-report.md
. Directory: 
tmp/reports/t3-{YYYY-MM-DD}-{case-id}/
.
Report structure (see references/report-schema.md):
  1. YAML metadata (wrapped in 
    ---
    ): For leaderboard parsing: 
    case_id
    , 
    source_url
    , 
    scores
    , 
    chart_data
    , 
    litmus_tests
    , 
    classification
    , etc.
  2. Text analysis: Aggregated from Auditors’ 
    extract_for_report
    : Product Overview, Tool Highlights, Toy Highlights, Trash Highlights, Final Judge Reasoning, Improvement Suggestions
Must include:
  • Basic product info (raw vs Brand-Blinded)
  • Triple Auditor score tables (format in auditor-templates.md)
  • Peer Review summary
  • Final Judge result
输出文件：固定为
99-audit-report.md
。存储目录：
tmp/reports/t3-{YYYY-MM-DD}-{case-id}/
。
报告结构（详见 references/report-schema.md）:
  1. YAML元数据（包裹在
    ---
    中）：用于排行榜解析，包含
    case_id
    、
    source_url
    、
    scores
    、
    chart_data
    、
    litmus_tests
    、
    classification
    等字段
  2. 文本分析：汇总自审计员的
    extract_for_report
    内容：产品概述、工具类亮点、玩具类亮点、垃圾类亮点、最终裁决推理、改进建议
必须包含的内容:
  • 基础产品信息（原始信息 vs 经过Brand Blinding处理的信息）
  • 三重审计员评分表（格式参考 auditor-templates.md）
  • 同行评审摘要
  • 最终裁决结果

Optional Branches

可选分支流程

  • Insufficient product info: Ask user for more description or references
  • Near boundary scores: Analyze dominant factors, explain classification
  • Auditor disagreement: Final Judge explains trade-off logic
  • Controversy: Offer multi-perspective analysis with evidence for each category
  • 产品信息不足：向用户索要更多描述或参考资料
  • 分数接近分类边界：分析主导因素，解释分类依据
  • 审计员意见分歧：最终裁决需解释权衡逻辑
  • 存在争议：提供多视角分析，并为每个分类提供证据

Resource Index

资源索引

  • Scripts (batch/offline):
    • scripts/crawl_product_info.py — Web crawling and product extraction; params: 
      --url <product_url> [--pretty] [--output <file>]
    • scripts/synthesize_results.py — Format Auditor results and compute classification; params: 
      --input <auditor_reports.json> [--pretty] [--output <file>]
  • References:
    • references/mandatory-page-list.md — Step 2, required; fixed crawl list
    • references/report-schema.md — Step 8, required; YAML schema
    • references/file-naming-convention.md — When creating report dirs; file naming
    • references/isolation-manifest-template.md — When creating 00-isolation-manifest.md
    • references/web-fetch-guide.md — When using web_fetch for product info
    • references/tool-auditor.md — Tool Auditor scoring
    • references/toy-auditor.md — Toy Auditor scoring
    • references/trash-auditor.md — Trash Auditor scoring
    • references/trash-red-flags.md — Trash high-sensitivity triggers
    • references/auditor-templates.md — Scoring tables; all Auditors must follow
    • references/t3-classification.md — Final Judge classification
    • references/scoring-checklists.md — Checklist index
    • references/defluff-guide.md — Brand Blinding
    • references/peer-review-guide.md — Peer Review
    • references/objective-data-standard.md — Objective data across stages
    • references/design-theories.md — Design theory background
  • 脚本（批量/离线使用）:
    • scripts/crawl_product_info.py — 网页爬取与产品提取；参数：
      --url <product_url> [--pretty] [--output <file>]
    • scripts/synthesize_results.py — 格式化审计员结果并计算分类；参数：
      --input <auditor_reports.json> [--pretty] [--output <file>]
  • 参考文档:
    • references/mandatory-page-list.md — 步骤2必填；固定爬取列表
    • references/report-schema.md — 步骤8必填；YAML schema
    • references/file-naming-convention.md — 创建报告目录时使用；文件命名规则
    • references/isolation-manifest-template.md — 创建00-isolation-manifest.md时使用
    • references/web-fetch-guide.md — 使用web_fetch获取产品信息时参考
    • references/tool-auditor.md — 工具类审计员评分指南
    • references/toy-auditor.md — 玩具类审计员评分指南
    • references/trash-auditor.md — 垃圾类审计员评分指南
    • references/trash-red-flags.md — 垃圾类高敏感度触发项
    • references/auditor-templates.md — 评分表；所有审计员必须遵循
    • references/t3-classification.md — 最终裁决分类规则
    • references/scoring-checklists.md — 评分清单索引
    • references/defluff-guide.md — Brand Blinding指南
    • references/peer-review-guide.md — 同行评审指南
    • references/objective-data-standard.md — 各阶段客观数据标准
    • references/design-theories.md — 设计理论背景

Important Notes

重要注意事项

  • 🚨 Information isolation (critical):
    • Auditors may only access Brand-Blinded info
    • Never raw product info, brand names, or marketing copy
    • Auditor tasks must state they evaluate based solely on Brand-Blinded info
    • Audit report must include "information source" field
  • 📊 Objective data completeness:
    • Collect complete objective data (specs, performance, reliability, market, sustainability, cost)
    • Brand Blinding must retain all objective data
    • Scores must be backed by verifiable objective data
    • See references/objective-data-standard.md
    • Report must include objective data summary and completeness score
  • Brand Blinding:
    • Must precede Auditor evaluation
    • Remove brand and marketing; keep objective data
  • Checklist-based scoring: Use reference checklists; each score must link to checklist items and evidence
  • Triple Auditor independence: Evaluate in parallel, no information sharing
  • Peer Review objectivity: Evidence-based; specific suggestions; open to feedback
  • Transparency: Process, reasoning, and evidence clearly documented
  • Agent-led: Brand Blinding, scoring, Peer Review, Final Judge run by agent; scripts only for crawl and formatting
  • 🚨 信息隔离（关键）:
    • 审计员仅能访问经过Brand Blinding处理的信息
    • 不得访问原始产品信息、品牌名称或营销文案
    • 审计员任务必须明确说明仅基于经过Brand Blinding处理的信息进行评估
    • 审计报告必须包含“信息来源”字段
  • 📊 客观数据完整性:
    • 收集完整的客观数据（参数、性能、可靠性、市场、可持续性、成本）
    • Brand Blinding必须保留所有客观数据
    • 评分必须有可验证的客观数据支持
    • 详见 references/objective-data-standard.md
    • 报告必须包含客观数据摘要及完整性评分
  • Brand Blinding:
    • 必须在审计员评估前执行
    • 移除品牌及营销内容；保留客观数据
  • 基于清单评分：使用参考清单；每一项评分必须关联清单条目及证据
  • 三重审计员独立性：并行评估，不得共享信息
  • 同行评审客观性：基于证据；给出具体建议；乐于接受反馈
  • 透明度：流程、推理过程及证据必须清晰记录
  • Agent主导：Brand Blinding、评分、同行评审、最终裁决均由Agent执行；脚本仅用于爬取及格式化

Examples

示例

Example 1: Full T3 Audit

示例1：完整T3审计

  • Flow: Full Brand Blinding + scoring + Peer Review + Final Judge
  • Execution: Agent crawls → Brand Blinding → parallel Auditors → Peer Review → optimization → Final Judge → report
  • Output: Full audit report with all sections
  • 流程：完整Brand Blinding + 评分 + 同行评审 + 最终裁决
  • 执行：Agent爬取 → Brand Blinding → 并行审计 → 同行评审 → 优化 → 最终裁决 → 生成报告
  • 输出：包含所有章节的完整审计报告

Example 2: Quick T3 Classification

示例2：快速T3分类

  • Flow: Simplified; focus on Brand Blinding and core scoring
  • Focus: Core value proposition, Litmus Tests, key checklist items
  • 流程：简化流程；重点关注Brand Blinding及核心评分
  • 评估重点：核心价值主张、测试标准、关键清单条目

Example 3: Competitor Audit

示例3：竞品审计

  • Flow: Batch crawl + parallel Auditors + Peer Review + comparison
  • Output: Comparison table + per-product reports
  • 流程：批量爬取 + 并行审计 + 同行评审 + 对比分析
  • 输出：对比表格 + 单产品报告

Audit Flow Diagram

审计流程图

User input URL
    ↓
Crawl product info (raw)
    ↓
Brand Blinding (Agent debrands)
    ├─ Raw info 🔒 (sealed)
    └─ Brand-Blinded info ✅ (only input for Auditors)
    ↓
┌─────────────────────────────────────────┐
│ Triple Auditor Scoring (Agent, parallel)│
│ 🔒 Isolation: Each Auditor only sees    │
│    Brand-Blinded info                   │
├───────────────┬─────────────┬────────────┤
│ 🟢 Tool       │ 🟡 Toy      │ 🔴 Trash   │
└───────┬───────┴─────┬───────┴──────┬────┘
        └─────────────┼──────────────┘
                      ↓
         Peer Review
                      ↓
         Auditor Optimization
                      ↓
         Final Judge
                      ↓
         99-audit-report.md
User input URL
    ↓
Crawl product info (raw)
    ↓
Brand Blinding (Agent debrands)
    ├─ Raw info 🔒 (sealed)
    └─ Brand-Blinded info ✅ (only input for Auditors)
    ↓
┌─────────────────────────────────────────┐
│ Triple Auditor Scoring (Agent, parallel)│
│ 🔒 Isolation: Each Auditor only sees    │
│    Brand-Blinded info                   │
├───────────────┬─────────────┬────────────┤
│ 🟢 Tool       │ 🟡 Toy      │ 🔴 Trash   │
└───────┬───────┴─────┬───────┴──────┬────┘
        └─────────────┼──────────────┘
                      ↓
         Peer Review
                      ↓
         Auditor Optimization
                      ↓
         Final Judge
                      ↓
         99-audit-report.md

🔒 Information Isolation

🔒 信息隔离规则

  1. Raw (Level 0): Brand names, marketing, emotional language — sealed, only Brand Blinding and Final Judge (for comparison)
  2. Brand-Blinded (Level 1): Functional and objective data — accessible to Auditors and Final Judge
Agent checklist when acting as Auditor:
  • Explicitly state use of Brand-Blinded info only
  • Base scores on reference checklists
  • No brand names referenced
  • No marketing language referenced
  • Report includes "Information source: Brand-Blinded product info"
  • Report includes "Scoring basis: [checklist name]"
  1. 原始信息（Level 0）：包含品牌名称、营销内容、情绪化语言 — 密封保存，仅用于Brand Blinding及最终裁决（用于对比）
  2. 经过Brand Blinding处理的信息（Level 1）：功能性及客观数据 — 可供审计员及最终裁决使用
Agent作为审计员时的检查清单：
  • 明确说明仅使用经过Brand Blinding处理的信息
  • 基于参考清单评分
  • 未引用任何品牌名称
  • 未引用任何营销话术
  • 报告包含“信息来源：经过Brand Blinding处理的产品信息”
  • 报告包含“评分依据：[清单名称]"