Performance Calibration Meeting Preparation / 绩效校准会议准备

> 中文使用说明：本技能用于绩效评估校准阶段，自动分析评级分布、识别边界案例、检测经理评分偏差，并生成完整校准会议数据包（HTML报告）。使用时请提供：员工初评名单（姓名+部门+经理+评级）、评级标准（如A/B+/B/B-/C）、强制分布目标（如A 10%, B+ 20%...）。可选提供：上期评级、KPI评分、自评、司龄等增强数据。直接在对话中粘贴数据即可触发。
Generate comprehensive data packages for performance calibration meetings, including rating distribution analysis, boundary case identification, manager bias detection, and meeting agenda generation.

When to Use

• User asks for 绩效校准 / 校准会议准备 / calibration prep
• Performance review cycle reaches the calibration stage (managers have submitted initial ratings)
• HR needs to prepare data packages for calibration committee meetings
• User mentions keywords: 绩效校准, 校准会议, 强制分布, 评分偏移, 绩效盘点, calibration

Do NOT Use When

• User wants to generate individual performance review reports (use performance-review-generator instead)
• User asks for 360-degree feedback aggregation (use feedback-360-aggregator instead)
• User wants to create PIP plans (use pip-generator instead)
• User is still in the goal-setting phase, not yet at rating stage

Input Requirements

Required Data

1. Employee list with initial ratings: Employee name/ID, department, manager, initial rating (e.g., A/B+/B/B-/C or 1-5 scale)
2. Rating scale definition: The company's performance rating scale and criteria
3. Forced distribution target (if applicable): The target percentage for each rating level

Optional Data (Enhances Analysis)

4. Prior period ratings: Same employees' ratings from last cycle — enables rating change tracking
5. Manager information: Manager name, department managed, number of direct reports
6. Employee metadata: Tenure, job level, role/position
7. Rating evidence: Self-assessment scores, peer review scores, KPI achievement rates
8. Compensation data: Current salary, last raise amount (for pay-for-performance check)

Data Input Formats

• Excel/CSV files with employee rating data
• Direct text/table input in conversation
• Structured data pasted from HR systems
• Partial data (skill will analyze what it can and note missing inputs)

Core Analysis Modules

Module 1: Rating Distribution Analysis (评级分布分析)

1.1 Company-Level Distribution

• Count and percentage of employees in each rating level
• Compare actual distribution vs forced distribution target
• Calculate variance (actual % - target %) for each level
• Flag levels where variance exceeds tolerance (default: ±5 percentage points)

1.2 Department-Level Distribution

• Rating distribution by department
• Cross-tabulation: Department x Rating Level
• Identify departments with significant deviation from company average
• Chi-square test (simplified): flag departments where distribution is statistically different

1.3 Distribution Visualization

• Stacked bar chart: actual vs target distribution
• Department comparison heatmap
• Rating funnel: employee count at each level

Module 2: Boundary Case Identification (边界案例识别)

Identify employees who are "on the fence" between two rating levels and need calibration discussion.

2.1 Score-Based Boundary Detection

If numeric scores are available:
- Define boundary zones: within ±5% of the cutoff score between two adjacent levels
- Example: If A/B+ cutoff is 90, then scores 85.5-94.5 are boundary cases
- Flag these cases as "Discuss" items

2.2 Qualitative Boundary Detection

If only letter ratings are available (no scores):
- Identify employees rated at the boundary levels (B+ and B- are most common boundary zones)
- Cross-reference with manager comments or self-assessment for discrepancy signals
- Flag employees where manager rating differs from self-assessment by 2+ levels

2.3 Boundary Case Classification

• Upgrade candidates: Currently at lower level, evidence supports upgrading
• Downgrade candidates: Currently at higher level, evidence supports downgrading
• Confirm candidates: Rating is appropriate but worth confirming in calibration
• Priority ranking: by impact (proximity to boundary, role criticality, salary implications)

Module 3: Manager Bias Detection (评分偏差检测)

3.1 Manager Leniency/Strictness Index

For each manager with 5+ direct reports:
- Leniency Index = (Actual average rating - Expected average rating) / Rating scale range
- Expected average = weighted average based on forced distribution
- Positive index = lenient manager (rates higher than expected)
- Negative index = strict manager (rates lower than expected)
- Severity thresholds:
- |Index| > 0.15: Severe bias (red flag)
- 0.08 < |Index| <= 0.15: Moderate bias (yellow flag)
- |Index| <= 0.08: Normal (green)

3.2 Central Tendency Detection

• Calculate the standard deviation of each manager's ratings
• Managers with low STD (< 0.5 on a 5-point scale) may be exhibiting central tendency bias
• Flag managers who cluster all ratings in the middle (e.g., everyone gets B)

3.3 Halo Effect Detection

• Compare each manager's ratings across different evaluation dimensions (if multi-dimensional data available)
• Low variance across dimensions for the same employee suggests halo effect
• Flag managers with average inter-dimension correlation > 0.85

3.4 Cross-Department Comparison

• Compare average ratings across departments of similar nature
• Identify outlier departments (significantly higher or lower than peer departments)
• Useful for inter-departmental calibration discussion

Module 4: Historical Trend Analysis (历史趋势分析)

When prior period data is available:

4.1 Rating Migration Matrix

Track how employees moved between rating levels:

         | To A | To B+ | To B | To B- | To C | Exit
From A   |  xx  |   xx  |  xx  |   xx  |  xx  |  xx
From B+  |  xx  |   xx  |  xx  |   xx  |  xx  |  xx
From B   |  xx  |   xx  |  xx  |   xx  |  xx  |  xx
From B-  |  xx  |   xx  |  xx  |   xx  |  xx  |  xx
From C   |  xx  |   xx  |  xx  |   xx  |  xx  |  xx

4.2 Rating Inflation/Deflation Detection

• Compare average rating this cycle vs last cycle
• If average rating rises while business performance is flat/declining -> rating inflation
• Calculate "Rating-to-Performance Gap": rating trend vs business KPI trend

4.3 Sticky Rating Detection

• Employees who have stayed at the same rating for 3+ consecutive cycles
• Not inherently bad, but worth reviewing: is it genuine consistency or rating inertia?
• Flag high-rating sticky cases (A or B+ for 3+ cycles) for validation
• Flag low-rating sticky cases (B- or C for 2+ cycles) for development discussion

Module 5: Pay-for-Performance Check (绩效薪酬一致性检查)

When compensation data is available:

5.1 Rating-Salary Correlation

• Average salary by rating level (should show ascending trend: A > B+ > B > B- > C)
• Flag inversions: employees with high ratings but low salary growth, or vice versa

5.2 Raise Distribution Analysis

• Average raise % by rating level
• Check if raise differentiation is meaningful (target: A gets 2-3x the raise of B)
• Identify "everyone gets similar raise" anti-pattern

Module 6: Meeting Agenda Generation (会议议程生成)

Based on all analyses above, generate a structured calibration meeting agenda:

6.1 Meeting Priority Queue

Prioritize discussion items:
1. P0 - Must discuss: Severe distribution deviation, manager bias red flags, boundary cases for critical roles
2. P1 - Should discuss: Moderate distribution deviation, boundary cases for non-critical roles, rating inflation signals
3. P2 - Information sharing: Normal cases, historical trends, department comparisons

6.2 Discussion Template per Case

For each boundary case to be discussed:
- Employee: Name, department, manager, current rating
- Rating evidence: KPI achievement, self-assessment, peer feedback summary
- Boundary reason: Why this case is flagged (score proximity, manager bias, distribution adjustment)
- Discussion question: Specific question for the calibration committee
- Suggested action: Upgrade / Downgrade / Confirm / Defer to next cycle

6.3 Time Allocation Guide

• Estimate discussion time per case (5-10 minutes for boundary cases, 2-3 minutes for confirmations)
• Total meeting time estimate
• Suggest discussion order (high-impact cases first)

Professional Logic: Calibration Decision Framework (校准决策框架)

Scenario A: Department Distribution Deviates from Target

部门分布偏离目标？
├── 偏离方向：高分过多（A/B+占比超出上限）
│   ├── 检查业务结果：部门KPI是否支持高分？
│   │   ├── KPI超额完成 → 可接受偏高，但需压缩边界案例
│   │   └── KPI未达标 → 要求经理下调评分，回归目标分布
│   └── 检查经理 leniency index
│       ├── 经理偏宽容 → 重点校准，压缩B+到B
│       └── 经理正常 → 可能是团队确实优秀，交叉验证
├── 偏离方向：低分过多（B-/C占比超出上限）
│   ├── 检查是否为新接手团队（磨合期影响）
│   ├── 检查是否有结构性问题（人员能力不足/管理问题）
│   └── 谨慎处理：避免一刀切下调，逐人讨论
└── 偏离方向：集中在中段（缺乏差异化）
    ├── 经理可能存在 central tendency bias
    ├── 要求经理重新区分：谁偏上、谁偏下
    └── 提供 KPI 数据辅助差异化判断

Scenario B: Manager Bias Detected

检测到经理评分偏差？
├── 宽容型偏差（Leniency Index > 0.15）
│   ├── 红旗标记，在校准会议中重点讨论其团队
│   ├── 要求经理提供每位高分员工的具体证据
│   └── 建议调整：将部分B+调整为B，将部分B调整为B-
├── 严格型偏差（Leniency Index < -0.15）
│   ├── 检查是否为高要求文化（可接受）还是标准过严
│   ├── 关注员工流失风险：严格评分是否导致离职？
│   └── 建议调整：将部分B-调整为B，将部分B调整为B+
└── 集中趋势偏差（STD < 0.5）
    ├── 经理未做充分区分
    ├── 要求经理基于 KPI 重新排序并强制分布
    └── 提供 calibration 训练

Scenario C: Rating Inflation Signal

评级通胀信号？
├── 定义：平均评级上升，但业务KPI持平或下降
├── 根因分析
│   ├── 经理护犊子：不愿给低分
│   ├── 标准漂移：今年评分标准比去年宽松
│   └── 团队变化：高绩效人员占比确实增加（需验证）
├── 校准措施
│   ├── 强制回归目标分布
│   ├── 引入相对排序（forced ranking within team）
│   └── 对比历史数据：同级别人员今年的评分是否比去年高
└── 沟通策略
    ├── 不公开否定经理评分，在校准会议上用数据引导
    └── 强调评分的激励资源约束（奖金池/调薪预算有限）

Scenario D: Employee Rating Dispute Potential

员工评级争议风险？
├── 高风险：自评与经理评分差异 ≥ 2级
│   ├── 校准会议中重点讨论
│   ├── 要求经理准备谈话要点和证据
│   └── HRBP 预备介入方案
├── 中风险：连续2期降级 或 连续3期同一评级
│   ├── 检查是否有客观依据
│   └── 准备发展计划或岗位调整建议
└── 低风险：评级变化在1级以内且有KPI支撑
    └── 正常走流程

Report Output Structure

Generate a visually appealing, self-contained HTML report with the following sections:

Section 1: Executive Summary (执行摘要)

• Total employees rated, number of departments, number of managers
• Overall distribution snapshot (one-line summary)
• Key issues identified (3-5 bullet points)
• Meeting priority: P0/P1/P2 case counts
• Display as a prominent hero banner

Section 2: Distribution Overview (分布总览)

• Company-level: actual vs target distribution table + bar chart
• Variance analysis with color coding (green/yellow/red)
• Key statistics: mean, median, STD of ratings

Section 3: Department Distribution Analysis (部门分布分析)

• Department x Rating cross-tabulation table
• Stacked bar chart showing each department's distribution
• Variance heatmap (departments x rating levels, colored by deviation magnitude)
• Outlier departments flagged with specific deviation details

Section 4: Manager Bias Detection (经理偏差检测)

• Manager leniency index ranking (bar chart with red/yellow/green markers)
• Central tendency detection results
• Specific managers flagged with details: team size, average rating, bias type, severity
• Recommended actions for flagged managers

Section 5: Boundary Cases (边界案例)

• Filterable list of employees near rating boundaries
• Each case shows: name, department, manager, current rating, boundary type, evidence summary
• Sorted by priority (upgrade candidates first, then downgrade, then confirm)
• Discussion question for each case
• Color-coded: green = upgrade candidate, yellow = confirm, red = downgrade

Section 6: Historical Trend (历史趋势)

• Rating migration matrix (if prior period data available)
• Average rating trend (this cycle vs last cycle)
• Rating inflation/deflation assessment
• Sticky rating cases flagged

Section 7: Meeting Agenda (会议议程)

• Time and date suggestion
• Participants list (calibration committee)
• Priority queue of discussion items (P0/P1/P2)
• Time allocation per item
• Total estimated meeting duration
• Discussion template for each P0 case

Section 8: Action Items Summary (行动项汇总)

• Post-meeting action checklist
• Rating adjustment tracking template
• Communication timeline (when to inform employees of final ratings)
• Appeals process reminder

HTML Design Guidelines

• Font: Use Noto Sans SC for Chinese text, loaded from Google Fonts
• Color theme: Professional performance review theme:
• Red (#e74c3c): Downgrade candidates, severe bias, P0 items
• Yellow (#f39c12): Boundary confirm cases, moderate bias, P1 items
• Green (#27ae60): Upgrade candidates, normal distribution, P2 items
• Blue (#3498db): Informational, neutral data
• Purple (#9b59b6): Agenda items, meeting structure
• Charts: Pure CSS visualizations:
• Distribution bars: CSS flexbox with colored segments
• Heatmap: CSS grid with background-color opacity
• Manager index: horizontal bar chart with gradient fills
• Migration matrix: CSS table with colored cells
• Layout: Responsive grid, max-width 1200px, mobile-friendly
• Cards: Rounded corners (16px), subtle shadows, colored top borders
• Self-contained: Single HTML file, no external dependencies except Google Fonts

Execution Steps

1. Gather Data: Collect initial ratings, rating scale, forced distribution target, and any optional data (prior ratings, manager info, compensation data).

2. Validate Data: Check for completeness, valid rating values, consistent naming. Flag missing data and proceed with available information.

3. Run Distribution Analysis: Calculate company-level and department-level distributions. Compare with forced distribution targets.

4. Identify Boundary Cases: Apply boundary detection logic. Classify into upgrade/downgrade/confirm categories with priority.

5. Detect Manager Bias: Calculate leniency index, central tendency, and halo effect for each manager. Flag problematic managers.

6. Historical Analysis: If prior data available, generate migration matrix and inflation assessment.

7. Pay-for-Performance Check: If compensation data available, run rating-salary correlation analysis.

8. Generate Meeting Agenda: Prioritize discussion items, allocate time, create discussion templates for P0 cases.

9. Build HTML Report: Generate complete self-contained HTML file with all 8 sections.

10. Save Report: Save to working directory as 绩效校准准备包_YYYYMMDD.html

11. Present Summary: Provide brief text summary of key findings and meeting preparation status.

Rating Scale Configuration

The skill should adapt to different rating scales. Common configurations:

5-Level Scale (most common in Chinese companies)

• A (卓越): Top 10-15%
• B+ (优秀): 15-25%
• B (良好): 40-50%
• B- (待改进): 15-20%
• C (不合格): 5-10%

4-Level Scale

• A (卓越): 15-20%
• B (良好): 30-40%
• C (合格): 30-40%
• D (不合格): 5-10%

3-Level Scale

• Exceeds (超额): 15-25%
• Meets (达标): 60-70%
• Below (未达标): 10-15%

Numeric Scale (1-5 or 1-100)

• Convert to distribution analysis by defining cutoff points
• Boundary zones calculated as ±5% of cutoff scores

If the user does not specify a rating scale or forced distribution, ask for it. If the user says "no forced distribution," skip target comparison and focus on relative analysis (department comparison, manager comparison, historical trends).

Example Usage

User provides:

请帮我准备绩效校准会议数据包，以下是Q2初评结果：

评级标准：5级（A/B+/B/B-/C）
强制分布：A 10%, B+ 20%, B 40%, B- 20%, C 10%

部门数据：
研发部（经理：张伟，25人）：
- A: 3人, B+: 7人, B: 12人, B-: 2人, C: 1人

销售部（经理：李娜，20人）：
- A: 1人, B+: 3人, B: 10人, B-: 4人, C: 2人

市场部（经理：王强，15人）：
- A: 2人, B+: 5人, B: 7人, B-: 1人, C: 0人

运营部（经理：刘洋，18人）：
- A: 4人, B+: 6人, B: 7人, B-: 1人, C: 0人

职能部（经理：陈静，12人）：
- A: 0人, B+: 2人, B: 8人, B-: 2人, C: 0人

上期（Q1）平均评级：3.2（B+偏下）
本期平均评级：3.35

Skill should:
1. Calculate total: 90人, A:10(11.1%), B+:23(25.6%), B:44(48.9%), B-:10(11.1%), C:3(3.3%)
2. Compare with target: A close, B+偏高(+5.6pp), B偏高(+8.9pp), B-偏低(-8.9pp), C偏低(-6.7%)
3. Flag: 运营部 A占比22.2% vs target 10% -> severe deviation (red)
4. Manager bias: 刘洋 leniency index high (too many A/B+), 陈静 central tendency (everyone B/B-)
5. Rating inflation: average rose from 3.2 to 3.35 (+4.7%), check if business KPIs support
6. Generate full calibration meeting data package