Back to Details

sales-lead-score

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Design a Lead Scoring Model

设计线索评分模型

Help the user design, weight, and tune a lead scoring model — from defining scoring dimensions and assigning point values through setting MQL/SQL thresholds and implementing in their tools. This skill is tool-agnostic and applies to any CRM (Salesforce, HubSpot), MAP (Marketo, Pardot), or enrichment provider (Apollo, ZoomInfo, Clearbit).
帮助用户设计、加权和优化线索评分模型——从定义评分维度、分配分值，到设置MQL/SQL阈值并在工具中落地。本Skill工具无关，适用于任何CRM（Salesforce、HubSpot）、MAP（Marketo、Pardot）或数据补全供应商（Apollo、ZoomInfo、Clearbit）。

Step 1 — Gather context

步骤1 — 收集背景信息

Ask the user:
  1. What do you sell?
    • Product/service category
    • Approximate ACV (affects scoring complexity — $5K ACV needs simpler scoring than $100K ACV)
  2. What's your sales motion?
    • A) Inbound-led — most leads come from marketing (content, ads, events)
    • B) Outbound-led — SDRs/BDRs source most leads
    • C) Product-led growth (PLG) — users sign up and self-serve, sales engages at usage thresholds
    • D) Hybrid — mix of inbound, outbound, and PLG
    • E) Channel/partner — leads come through partners
  3. What tools do you use?
    • CRM: Salesforce, HubSpot, Pipedrive, other
    • MAP: Marketo, HubSpot, Pardot, ActiveCampaign, other
    • Enrichment: Apollo, ZoomInfo, Clearbit, other
    • Product analytics (for PLG): Amplitude, Mixpanel, Heap, Segment, other
  4. Current scoring situation?
    • A) Starting from scratch — no scoring model exists
    • B) Have a model but it's not working (describe symptoms)
    • C) Have a basic model, want to improve it
    • D) Rebuilding after ICP change or new product launch
  5. What does your funnel look like today?
    • Monthly lead volume (rough)
    • Current MQL → SQL conversion rate (if known)
    • Current SQL → closed-won rate (if known)
If the user's request already provides most of this context, skip directly to the relevant step. Lead with your best-effort answer using reasonable assumptions (stated explicitly), then ask only the most critical 1-2 clarifying questions at the end — don't gate your response behind gathering complete context.
询问用户：
  1. 你销售的是什么？
    • 产品/服务类别
    • 大致ACV（年度合同价值，影响评分模型复杂度——ACV为5000美元的业务所需的评分模型比ACV为10万美元的更简单）
  2. 你的销售模式是什么？
    • A) inbound主导——大部分线索来自营销（内容、广告、活动）
    • B) outbound主导——SDR/BDR开发大部分线索
    • C) 产品驱动增长（PLG）——用户自主注册使用，销售在达到特定使用阈值时介入
    • D) 混合模式——inbound、outbound和PLG结合
    • E) 渠道/合作伙伴——线索来自合作伙伴
  3. 你使用哪些工具？
    • CRM：Salesforce、HubSpot、Pipedrive或其他
    • MAP：Marketo、HubSpot、Pardot、ActiveCampaign或其他
    • 数据补全：Apollo、ZoomInfo、Clearbit或其他
    • 产品分析（PLG场景）：Amplitude、Mixpanel、Heap、Segment或其他
  4. 当前评分模型的情况？
    • A) 从零开始——尚无评分模型
    • B) 已有模型但效果不佳（请描述症状）
    • C) 已有基础模型，想要优化
    • D) 客户画像（ICP）变更或新产品发布后重建模型
  5. 当前销售漏斗的情况？
    • 月度线索量（大致数据）
    • 当前MQL→SQL转化率（若已知）
    • 当前SQL→成单率（若已知）
如果用户的请求已包含大部分上述信息，可直接跳至对应步骤。 先基于合理假设提供初步方案（需明确说明假设），最后仅询问最关键的1-2个澄清问题——不要因要求收集完整背景而延迟给出回复。

Step 2 — Define scoring dimensions

步骤2 — 定义评分维度

Build a scoring model across four dimensions. Default weights are a starting point — tune based on your sales motion.
从四个维度构建评分模型。默认权重为起始参考值，可根据销售模式调整。

Dimension 1: Demographic fit (default 25%)

维度1：人口统计匹配度（默认25%）

Score how well the individual matches your buyer persona.
AttributeHigh scoreMedium scoreLow/negative score
Job titleExact ICP title match (e.g., VP Engineering)Adjacent title (Director of Engineering, Head of Platform)Unrelated title (HR Manager when selling to Engineering)
SeniorityDecision-maker level for your productInfluencer levelToo junior to buy or influence
DepartmentPrimary buying departmentAdjacent departmentUnrelated department
Job functionDirect match to problem you solveRelated functionNo relevance
Example point values (25 points max):
  • Title exact match: 10 pts | Adjacent: 5 pts | No match: 0 pts
  • Seniority match: 8 pts | Adjacent: 4 pts | Too junior: 0 pts
  • Department match: 7 pts | Adjacent: 3 pts | Unrelated: 0 pts
评估线索个人与你的客户画像（ICP）的匹配程度。
属性高分标准中分标准低分/负分标准
职位头衔完全匹配ICP目标头衔（如VP Engineering）相近头衔（Director of Engineering、Head of Platform）无关头衔（销售给技术部门时的HR经理）
职级产品的决策层级影响者层级职级过低，无购买或影响力
部门核心采购部门相关部门无关部门
职能直接匹配产品解决的问题场景相关职能无关联
示例分值（最高25分）：
  • 头衔完全匹配：10分 | 相近匹配：5分 | 不匹配：0分
  • 职级匹配：8分 | 相近匹配：4分 | 职级过低：0分
  • 部门匹配：7分 | 相近匹配：3分 | 无关：0分

Dimension 2: Firmographic fit (default 25%)

维度2：企业特征匹配度（默认25%）

Score how well the company matches your ICP.
AttributeHigh scoreMedium scoreLow/negative score
Company sizeSweet spot (e.g., 100-500 employees)Adjacent range (50-100 or 500-1000)Way outside range
IndustryPrimary target industryAdjacent industryIndustry you don't serve
RevenueRevenue range that matches your pricingAdjacent rangeCan't afford your product
GeographyPrimary marketServiceable marketUnsupported region
Tech stackUses complementary technologyNeutral tech stackUses competing product (could be positive for displacement)
Example point values (25 points max):
  • Company size sweet spot: 8 pts | Adjacent: 4 pts | Outside: 0 pts
  • Industry match: 7 pts | Adjacent: 3 pts | No match: 0 pts
  • Revenue fit: 5 pts | Adjacent: 2 pts | Outside: 0 pts
  • Geography: 3 pts | Serviceable: 1 pt | Unsupported: 0 pts
  • Tech stack fit: 2 pts
评估企业与你的客户画像（ICP）的匹配程度。
属性高分标准中分标准低分/负分标准
企业规模理想区间（如100-500名员工）相近区间（50-100或500-1000名）远超出目标区间
行业核心目标行业相关行业你不服务的行业
营收与定价匹配的营收区间相近区间无法承担你的产品
地域核心市场可服务市场未覆盖区域
技术栈使用互补技术中立技术栈使用竞品（可能是置换机会）
示例分值（最高25分）：
  • 企业规模理想区间：8分 | 相近区间：4分 | 超出区间：0分
  • 行业匹配：7分 | 相关行业：3分 | 不匹配：0分
  • 营收匹配：5分 | 相近区间：2分 | 超出区间：0分
  • 地域：3分 | 可服务市场：1分 | 未覆盖区域：0分
  • 技术栈匹配：2分

Dimension 3: Behavioral signals (default 30%)

维度3：行为信号（默认30%）

Score what the lead is doing — this is the most predictive dimension for most teams.
SignalPointsDecay
Requested demo/trial15 ptsNone — this is a hard conversion event
Pricing page visit10 ptsDecays to 5 after 14 days
Multiple website visits (3+ in 7 days)8 ptsDecays to 4 after 14 days
Content download (ebook, whitepaper)5 ptsDecays to 2 after 30 days
Email engagement (open + click)3 pts per engagementDecays to 0 after 30 days
Webinar/event attendance8 ptsDecays to 4 after 30 days
Intent data — researching your category10 ptsDecays to 5 after 14 days (intent is perishable)
G2/review site comparison views8 ptsDecays to 4 after 14 days
For PLG/product-led motions, add product usage signals:
SignalPointsDecay
Signed up for free tier/trial10 ptsNone
Completed onboarding8 ptsNone
Hit usage threshold (e.g., 100 API calls, 5 team members)15 ptsNone
Invited team members10 ptsNone
Used premium feature (paywall hit)12 ptsDecays to 6 after 30 days
Daily active usage (5+ days in last 14)10 ptsRolling — recalculated weekly
评估线索的行为——这是对大多数团队最具预测性的维度。
行为信号分值分数衰减
申请演示/试用15分无衰减——这是明确的转化事件
访问定价页面10分14天后衰减至5分
多次访问网站（7天内3次以上）8分14天后衰减至4分
下载内容（电子书、白皮书）5分30天后衰减至2分
邮件互动（打开+点击）每次3分30天后衰减至0分
参加网络研讨会/活动8分30天后衰减至4分
意向数据——研究你的产品类别10分14天后衰减至5分（意向具有时效性）
查看G2/评测网站的竞品对比8分14天后衰减至4分
针对PLG/产品驱动模式，需添加产品使用信号：
行为信号分值分数衰减
注册免费版/试用10分无衰减
完成新手引导8分无衰减
达到使用阈值（如100次API调用、5名团队成员）15分无衰减
邀请团队成员10分无衰减
使用付费功能（触发付费墙）12分30天后衰减至6分
每日活跃使用（过去14天内5天以上）10分滚动计算——每周重新统计

Dimension 4: Timing signals (default 20%)

维度4：时机信号（默认20%）

Score recency and urgency signals.
SignalPointsDecay
New in role (<90 days)10 ptsDecays to 5 after 90 days, 0 after 180
Recent funding8 ptsDecays to 4 after 90 days
Hiring for roles your product supports6 ptsDecays to 3 after 30 days (job postings are time-sensitive)
Company growth (20%+ headcount in 6 months)5 ptsDecays to 2 after 90 days
Competitor contract renewal window10 ptsDecays to 0 after the window passes
评估时效性和紧迫性信号。
信号分值分数衰减
新入职（90天内）10分90天后衰减至5分，180天后衰减至0分
近期融资8分90天后衰减至4分
招聘与你的产品相关的岗位6分30天后衰减至3分（招聘信息时效性强）
企业增长（6个月内员工增长20%以上）5分90天后衰减至2分
竞品合同续约窗口期10分窗口期结束后衰减至0分

Tuning weights by sales motion

根据销售模式调整权重

MotionDemographicFirmographicBehavioralTiming
Inbound-led20%20%40%20%
Outbound-led25%30%20%25%
PLG15%15%50%20%
Enterprise/ABM25%25%25%25%
销售模式人口统计企业特征行为信号时机信号
Inbound主导20%20%40%20%
Outbound主导25%30%20%25%
PLG模式15%15%50%20%
企业/ABM模式25%25%25%25%

Step 3 — Set thresholds & stages

步骤3 — 设置阈值与阶段

MQL/SQL threshold calibration

MQL/SQL阈值校准

Start with these defaults, then tune based on conversion data:
ThresholdDefaultWhat it triggers
MQL (Marketing Qualified Lead)Top 20% of scored leadsMarketing nurture intensifies, SDR notification
SQL (Sales Qualified Lead)Top 5% of scored leadsSDR outreach, AE handoff, or sales follow-up
PQL (Product Qualified Lead, PLG only)Usage threshold + firmographic fitSales outreach to active free users
How to set initial thresholds:
  1. Score your last 100 closed-won deals retroactively
  2. Find the median score — this is roughly your SQL threshold
  3. Set MQL at 60-70% of the SQL threshold
  4. Adjust after 30 days of live data
从默认值开始，再根据转化数据调整：
阈值默认值触发动作
MQL（营销合格线索）评分排名前20%的线索强化营销培育、通知SDR
SQL（销售合格线索）评分排名前5%的线索SDR跟进、对接销售代表或触发销售跟进
PQL（产品合格线索，仅PLG模式）使用阈值+企业特征匹配向活跃免费用户发起销售跟进
初始阈值设置方法：
  1. 回溯评分最近100个成单客户
  2. 找到中位分值——这大致是你的SQL阈值
  3. 将MQL阈值设置为SQL阈值的60-70%
  4. 上线30天后根据实际数据调整

Scoring decay rules

评分衰减规则

Behavioral signals lose relevance over time. Implement decay to prevent score inflation:
  • Fast decay (7-14 days): Intent data, pricing page visits, comparison shopping
  • Medium decay (30 days): Content downloads, email engagement, webinar attendance
  • Slow decay (90 days): Job changes, funding events, hiring signals
  • No decay: Demo requests, trial signups, firmographic/demographic fit
行为信号的相关性会随时间降低。实施衰减规则以避免分数虚高：
  • 快速衰减（7-14天）：意向数据、定价页面访问、竞品对比浏览
  • 中等衰减（30天）：内容下载、邮件互动、网络研讨会参与
  • 慢速衰减（90天）：职位变更、融资事件、招聘信号
  • 无衰减：演示申请、试用注册、企业/人口统计匹配度

Negative scoring

负向评分

Subtract points for disqualifying signals:
SignalPoints
Unsubscribed from emails-20 pts
Competitor employee-50 pts (or auto-disqualify)
Student/educational email (.edu)-30 pts
Personal email (gmail, yahoo) for B2B product-10 pts
Job title contains "intern" or "student"-20 pts
Company size way below minimum-15 pts
Bounced email-10 pts
Marked as "do not contact"Auto-disqualify
对不合格线索扣除分数：
信号扣分
退订邮件-20分
竞品员工-50分（或直接 disqualify）
学生/教育邮箱（.edu）-30分
B2B产品使用个人邮箱（gmail、yahoo）-10分
职位头衔含“intern”或“student”-20分
企业规模远低于最小值-15分
邮件退回-10分
标记为“请勿联系”直接 disqualify

Step 4 — Implementation guide

步骤4 — 落地指南

HubSpot

HubSpot

  1. Properties > Create a custom "Lead Score" number property
  2. Workflows > Create scoring workflows that add/subtract points based on triggers
  3. For behavioral scoring, use "Contact activity" triggers (page views, form fills, email clicks)
  4. For decay, create time-based workflows that reduce scores after X days
  5. Set MQL lifecycle stage change when score exceeds threshold
  1. 进入Properties > 创建自定义“Lead Score”数值属性
  2. 进入Workflows > 创建评分工作流，根据触发条件增减分数
  3. 行为评分使用“Contact activity”触发器（页面访问、表单提交、邮件点击）
  4. 分数衰减：创建时间触发的工作流，X天后降低分数
  5. 设置分数超过阈值时自动变更MQL生命周期阶段

Salesforce

Salesforce

  1. Setup > Lead Scoring (if using Einstein Lead Scoring) or custom fields + Process Builder/Flow
  2. Create a "Lead Score" number field on Lead and Contact objects
  3. Use Flow Builder to increment/decrement based on field changes and activities
  4. For behavioral scoring, integrate with your MAP (Marketo, Pardot) which tracks engagement
  5. Create assignment rules that route SQLs to the right rep
  1. 进入Setup > Lead Scoring（若使用Einstein Lead Scoring）或自定义字段 + Process Builder/Flow
  2. 在Lead和Contact对象上创建“Lead Score”数值字段
  3. 使用Flow Builder根据字段变更和行为增减分数
  4. 行为评分需集成MAP（Marketo、Pardot）以追踪互动数据
  5. 创建分配规则，将SQL路由给对应销售代表

Marketo

Marketo

  1. Admin > Scoring > New Scoring Model
  2. Define demographic score (person attributes) and behavioral score (activities) separately
  3. Set up smart campaigns that add/subtract points for each trigger
  4. Configure scoring decay with scheduled batch campaigns
  5. Sync score to Salesforce for routing and visibility
  1. 进入Admin > Scoring > New Scoring Model
  2. 分别定义人口统计评分（个人属性）和行为评分（行为活动）
  3. 创建智能营销活动，根据触发条件增减分数
  4. 通过定时批量活动配置评分衰减
  5. 将分数同步至Salesforce以实现路由和可视化

Apollo

Apollo

  1. Use Apollo's built-in lead scoring (Settings > Scoring) for basic ICP fit scoring
  2. For behavioral scoring, supplement with your MAP — Apollo's native scoring is primarily demographic/firmographic
  3. Use Apollo's intent data as an input to behavioral scoring in your MAP
  4. Export scored leads to your CRM for routing
  1. 使用Apollo内置的线索评分（Settings > Scoring）进行基础ICP匹配评分
  2. 行为评分需结合你的MAP补充——Apollo原生评分主要针对人口/企业统计特征
  3. 将Apollo的意向数据作为MAP中行为评分的输入
  4. 将评分后的线索导出至CRM进行路由

Testing the model

模型测试

Before going live:
  1. Backtest: Score your last 6 months of leads retroactively. Check: do closed-won deals score higher than closed-lost?
  2. Threshold check: At your proposed MQL threshold, what % of leads qualify? (Target: 15-25%)
  3. False positive check: Sample 20 leads above MQL threshold — would a rep actually want to call them?
  4. False negative check: Sample 20 closed-won deals — did they score above MQL before they closed?
上线前需完成：
  1. 回溯测试：对过去6个月的线索进行回溯评分。检查：成单客户的分数是否高于未成交客户？
  2. 阈值检查：按你设定的MQL阈值，有多少比例的线索符合条件？（目标：15-25%）
  3. 误报检查：抽样20个MQL以上的线索——销售代表是否真的愿意联系他们？
  4. 漏报检查：抽样20个成单客户——他们在成单前的分数是否达到MQL标准？

Step 5 — Tuning & maintenance

步骤5 — 优化与维护

Monthly review cadence

月度回顾节奏

Every month, review:
  1. Conversion rates by score band: Are high-score leads actually converting better?
  2. MQL → SQL conversion: If below 30%, your MQL threshold is too low (too many bad MQLs)
  3. SQL → opportunity rate: If below 50%, your SQL threshold is too low
  4. Score distribution: Is the curve healthy or are most leads clustered at one score?
  5. Sales feedback: Are reps ignoring MQLs? If so, the model isn't surfacing quality leads.
每月需回顾：
  1. 各分数段的转化率：高分线索的转化率是否确实更高？
  2. MQL→SQL转化率：若低于30%，说明MQL阈值过低（无效MQL过多）
  3. SQL→商机率：若低于50%，说明SQL阈值过低
  4. 分数分布：分数曲线是否健康？还是大部分线索集中在同一分数段？
  5. 销售反馈：销售是否忽略MQL？若是，说明模型未筛选出高质量线索

Weight adjustment methodology

权重调整方法

  1. Export all leads from last 90 days with their scores and outcomes (converted vs not)
  2. For each scoring dimension, calculate conversion rate by score band
  3. If a dimension has no correlation with conversion → reduce its weight
  4. If a dimension strongly correlates → increase its weight
  5. Re-normalize so total weights = 100%
  1. 导出过去90天的所有线索及其分数和转化结果（成交/未成交）
  2. 针对每个评分维度，计算各分数段的转化率
  3. 若某维度与转化率无相关性——降低其权重
  4. 若某维度与转化率强相关——增加其权重
  5. 重新归一化，确保总权重为100%

Common failure modes

常见失效场景

SymptomCauseFix
Too many MQLs, sales ignores themMQL threshold too lowRaise threshold, add more behavioral weight
Too few MQLs, pipeline starvingThreshold too high or scoring too restrictiveLower threshold, check if firmographic filters are too narrow
High-score leads don't convertDemographic fit overweighted, behavioral underweightedIncrease behavioral weight, add decay to stale signals
Score inflation over timeNo decay rules, points only go upImplement decay on all behavioral signals
Model works for inbound, not outboundModel only has behavioral signalsAdd firmographic and timing dimensions
Sales and marketing disagree on MQL definitionModel built without sales inputCo-create thresholds with sales leadership, review monthly
症状原因解决方案
MQL过多，销售忽略MQL阈值过低提高阈值，增加行为信号权重
MQL过少，漏斗缺乏线索阈值过高或评分规则过严降低阈值，检查企业特征筛选是否过窄
高分线索未成交人口统计匹配度过重，行为信号权重不足增加行为信号权重，为陈旧信号添加衰减规则
分数虚高无衰减规则，分数只增不减在所有行为信号上实施衰减规则
模型对inbound有效，但对outbound无效模型仅包含行为信号添加企业特征和时机维度
销售与营销对MQL定义有分歧模型未结合销售意见构建与销售领导层共同制定阈值，每月联合回顾

Quarterly recalibration

季度重新校准

Every quarter:
  1. Re-run the backtest with latest data
  2. Check if ICP has shifted (new verticals, new personas)
  3. Review and update point values based on conversion data
  4. Add/remove scoring signals based on new data sources
  5. Validate thresholds still align with funnel capacity
每季度需完成：
  1. 使用最新数据重新进行回溯测试
  2. 检查客户画像（ICP）是否变更（新垂直领域、新目标人群）
  3. 根据转化数据回顾并更新分值
  4. 根据新数据源添加/移除评分信号
  5. 验证阈值是否仍与漏斗容量匹配

Gotchas

注意事项

  • Don't weight demographics too heavily. Claude defaults to giving 50%+ weight to title/seniority because it's easy to match. But behavioral signals (what they're doing) are more predictive than demographics (who they are). Start with at least 30% behavioral weight.
  • Don't skip negative scoring. A lead can have a perfect title at a perfect company but be a student, a competitor, or already unsubscribed. Negative scores prevent false positives that waste sales time and damage your credibility with the sales team.
  • Don't set static thresholds and forget them. Scoring models drift as your ICP evolves and market conditions change. A model that was calibrated 6 months ago may be surfacing the wrong leads today. Review and recalibrate quarterly using actual conversion data.
  • Don't build the model in isolation. Sales and marketing must agree on MQL/SQL definitions. A scoring model that marketing builds without sales input leads to "bad MQLs" complaints and erodes trust. Co-create with sales leadership and review together monthly.
  • 不要过度加权人口统计特征：Claude默认会给头衔/职级50%以上的权重，因为容易匹配。但行为信号（线索的行为）比人口统计特征（线索是谁）更具预测性。起始时行为信号权重至少设置为30%。
  • 不要跳过负向评分：线索可能拥有完美的头衔和企业背景，但可能是学生、竞品员工或已退订邮件。负向评分可避免浪费销售时间，同时维护销售团队对你的信任。
  • 不要设置静态阈值后就置之不理：随着客户画像（ICP）演变和市场环境变化，评分模型会逐渐失效。6个月前校准的模型如今可能筛选出错误的线索。需每季度使用实际转化数据回顾和重新校准。
  • 不要孤立构建模型：销售和营销必须就MQL/SQL定义达成一致。营销单方面构建的评分模型会导致“无效MQL”的抱怨，损害双方信任。需与销售领导层共同创建阈值，每月联合回顾。

Related skills

相关Skill

  • /sales-intent
    — Read buying signals that feed into your scoring model
  • /sales-prospect-list
    — Build prospect lists to score
  • /sales-enrich
    — Enrich leads with the demographic/firmographic data you need to score
  • /revops
    — Design the broader marketing-to-sales handoff process around your scoring model
  • /sales-apollo
    — Set up Apollo's native scoring features
  • /sales-do
    — Not sure which skill to use? The router matches any sales objective to the right skill. Install: 
    npx skills add sales-skills/sales --skills sales-do
  • /sales-intent
    — 读取可输入到评分模型的购买信号
  • /sales-prospect-list
    — 构建待评分的潜在客户列表
  • /sales-enrich
    — 补充评分所需的人口/企业统计数据
  • /revops
    — 围绕评分模型设计更广泛的营销到销售交接流程
  • /sales-apollo
    — 配置Apollo的原生评分功能
  • /sales-do
    — 不确定使用哪个Skill？该路由工具可将任何销售目标匹配到对应Skill。安装：
    npx skills add sales-skills/sales --skills sales-do

Examples

示例

Example 1: Inbound B2B SaaS scoring model

示例1：Inbound主导的B2B SaaS评分模型

User says: "Help me build a lead scoring model for our B2B SaaS product. We're inbound-heavy, $50K ACV." Skill does:
  1. Designs a 4-dimension model weighted for inbound (40% behavioral, 20% demographic, 20% firmographic, 20% timing)
  2. Assigns point values for each attribute with specific thresholds
  3. Sets MQL at ~60 points, SQL at ~85 points
  4. Provides HubSpot/Salesforce implementation steps
  5. Creates a testing and tuning plan Result: Full scoring model with calibrated thresholds, implementation guide, and monthly review cadence
用户需求：“帮我为我们的B2B SaaS产品构建线索评分模型。我们以inbound为主，ACV为5万美元。” Skill执行：
  1. 设计针对inbound模式的四维模型（行为信号40%、人口统计20%、企业特征20%、时机信号20%）
  2. 为每个属性分配具体分值和阈值
  3. 设置MQL阈值约60分，SQL阈值约85分
  4. 提供HubSpot/Salesforce落地步骤
  5. 制定测试和优化计划 结果：包含校准阈值、落地指南和月度回顾节奏的完整评分模型

Example 2: Fixing a broken scoring model

示例2：修复失效的评分模型

User says: "Our MQL-to-SQL conversion rate is 8%. Our scoring model isn't working." Skill does:
  1. Diagnoses likely causes (threshold too low, demographic overweighting, no decay)
  2. Recommends scoring audit — backtest recent leads against outcomes
  3. Proposes weight adjustments (increase behavioral, decrease demographic)
  4. Suggests threshold recalibration based on conversion data
  5. Creates a 30-day improvement plan with specific metrics to track Result: Diagnosis of scoring model issues with specific fixes and a recalibration plan
用户需求：“我们的MQL→SQL转化率只有8%。评分模型不起作用。” Skill执行：
  1. 诊断可能原因（阈值过低、人口统计权重过高、无衰减规则）
  2. 建议进行评分审计——对近期线索进行回溯评分并对比结果
  3. 提出权重调整方案（增加行为信号权重，降低人口统计权重）
  4. 建议根据转化数据重新校准阈值
  5. 制定30天改进计划及具体跟踪指标 结果：评分模型问题诊断、具体修复方案和重新校准计划

Example 3: PLG product usage scoring

示例3：PLG产品使用评分

User says: "We're a PLG company. How should product usage signals factor into lead scoring?" Skill does:
  1. Designs a behavioral-heavy model (50% behavioral including product signals)
  2. Defines PQL criteria based on usage thresholds (API calls, team members, feature adoption)
  3. Integrates product analytics events as scoring signals
  4. Sets PQL threshold separately from MQL/SQL
  5. Designs handoff trigger from product usage → sales outreach Result: PLG-adapted scoring model with product event scoring, PQL definition, and sales handoff triggers
用户需求：“我们是PLG公司。产品使用信号应如何融入线索评分？” Skill执行：
  1. 设计以行为信号为主的模型（50%行为信号，含产品使用数据）
  2. 基于使用阈值（API调用、团队成员、功能采用率）定义PQL标准
  3. 将产品分析事件集成作为评分信号
  4. 为PQL设置独立于MQL/SQL的阈值
  5. 设计从产品使用到销售跟进的触发机制 结果：适配PLG模式的评分模型，包含产品事件评分、PQL定义和销售跟进触发规则

Troubleshooting

故障排除

Sales says "these MQLs are garbage"

销售反馈“这些MQL毫无价值”

Cause: MQL threshold is too low, or the model overweights demographic fit without behavioral validation Solution: Review the last 50 MQLs that sales rejected. Look for patterns — are they the wrong persona? Right persona but not engaged? Adjust the dimension that's causing false positives. Often the fix is adding a minimum behavioral score requirement on top of the overall threshold.
原因：MQL阈值过低，或模型过度加权人口统计特征而未结合行为验证 解决方案：回顾销售拒绝的最近50个MQL，寻找规律——是客户画像错误？还是画像正确但线索未产生互动？调整导致误报的维度。通常的解决方法是在总分阈值基础上添加最低行为分数要求。

Score inflation — everyone is an MQL

分数虚高——所有人都是MQL

Cause: No decay rules, points only accumulate, never decrease Solution: Implement decay on all behavioral signals (14-day for intent, 30-day for content, 90-day for timing). Run a one-time score recalculation after implementing decay.
原因：无衰减规则，分数只增不减 解决方案：对所有行为信号实施衰减规则（意向数据14天、内容30天、时机信号90天）。实施衰减后运行一次全量分数重算。

Model performs differently for different segments

模型在不同细分群体中的表现差异大

Cause: One-size-fits-all model doesn't account for segment-specific buying patterns Solution: Consider separate scoring models for distinct segments (enterprise vs SMB, inbound vs outbound). At minimum, adjust firmographic fit scoring to not penalize enterprise leads for different engagement patterns (they visit fewer pages but have higher deal sizes).
原因：通用模型未考虑细分群体的特定购买模式 解决方案：考虑为不同细分群体（企业vs SMB、inbound vs outbound）设置独立的评分模型。至少调整企业特征评分规则，避免因企业的不同互动模式（如企业客户访问页面更少但 deal size更大）而 penalize他们。