Agent Skills
› mohitagw15856/pm-claude-skills
› metric-gaslighting-detector
metric-gaslighting-detector
GitHub审计仪表板或KPI报告中的叙事误导，识别分母游戏、幸存者偏差等11种数据扭曲。输出包含失真评分表、诚实重述、向负责人提问的三个关键问题及建议图表，揭示数据背后的真实结论。
Trigger Scenarios
发现数据过于整洁或叙事依赖单一图表
继承未定义指标需验证其可靠性
怀疑KPI报告存在误导性排列
Install
npx skills add mohitagw15856/pm-claude-skills --skill metric-gaslighting-detector -g -y
SKILL.md
Frontmatter
{
"name": "metric-gaslighting-detector",
"description": "Find out how a dashboard, KPI report, or metrics slide is lying to you — before you repeat its story in a bigger room. Use when numbers feel too tidy, a narrative rests on one chart, or you inherited metrics you didn't define. Produces a deception audit: every metric graded for the eleven classic distortions (denominator games, survivorship, y-axis crimes, cherry-picked windows…), the story the data would tell under honest framing, and the three questions to ask the metric's owner."
}
Metric Gaslighting Detector
Dashboards rarely contain false numbers. They contain true numbers arranged to create false beliefs. This skill audits the arrangement — the eleven standard distortions through which honest data becomes dishonest narrative.
Required Inputs
- The metrics artifact — the dashboard description, KPI table, chart, or the numbers with their labels exactly as presented. Include axis ranges, time windows, and any annotations; the lie usually lives there.
- The claim being made with it (if any) — "churn is under control", "the launch worked". The audit tests the claim-data connection, not the data alone.
The Eleven Distortions
- Denominator games — the base changed ("of active users" quietly became "of weekly active")
- Survivorship framing — measuring only what remained (retention of cohorts that didn't churn early)
- Y-axis crimes — truncated baselines, dual axes, log scales without labels
- The cherry window — the date range that starts at the trough or ends before the drop
- Mix-shift laundering — the aggregate improved because composition changed, not performance
- Ratio without magnitude — "+40%!" concealing 5→7
- The vanity proxy — measuring what moves instead of what matters (signups for activation)
- Goodhart's ghost — the metric improved because it became a target, and the gamed behaviour is visible elsewhere
- Smoothing to silence — rolling averages wide enough to bury the event being asked about
- The missing counterfactual — "up 20% since launch" with no baseline trend (it was up 25% before)
- Significance theatre — differences within noise presented as movement ("ticked up to 4.6 from 4.5, n=41")
Output Format
- The audit table — metric | distortion(s) detected | severity (🔴 changes the conclusion / 🟡 shades it / 🟢 clean) | the honest version of that number's sentence.
- The honest retelling (≤150 words) — what this data says under fair framing. Sometimes the story survives; say so — the detector earns trust by clearing metrics too.
- Three questions for the owner — specific, answerable, non-accusatory ("what was the trend in the 8 weeks before launch?"), ordered by how much the answer would change the conclusion.
- The one chart to request — the single re-cut (full window, fixed denominator, split by segment) that would settle the biggest 🔴.
Quality Checks
- Every 🔴 names the specific mechanism and what the conclusion becomes without it — "misleading" alone is not a finding
- At least one metric is graded 🟢 or the audit admits the artifact gave nothing to clear — all-guilty audits read as motivated
- The honest retelling uses only the numbers present — the detector doesn't smuggle in its own speculation
- Questions are answerable from data the owner plausibly has, and none contain an accusation
- Distortion names from the list are used consistently so repeated audits build a shared vocabulary
Anti-Patterns
- Do not accuse people of lying — the framing is "what belief does this arrangement create vs what the data supports"; most gaslighting dashboards are self-deception forwarded
- Do not grade a metric 🔴 for a distortion that doesn't change the decision at hand — severity is about consequences, not purity
- Do not demand data that doesn't exist as a gotcha — the three questions must be realistically answerable
- Do not rewrite the numbers — the honest retelling reframes; it never adjusts figures
- Do not skip auditing metrics that support conclusions you like — run the eleven on the favourable ones first
Version History
- 961cbeb Current 2026-07-11 19:48
