Agent Skills
› mohitagw15856/pm-claude-skills
› tornado-sensitivity
tornado-sensitivity
GitHub用于对模型输出进行单变量敏感性分析，生成按影响排序的龙卷风图。识别关键驱动因素，分配尽调精力，避免无效争论，并提供Excel报告及会议结论。
触发场景
需要对LTV、ROI或预测等模型输出进行敏感性分析
团队在争论非关键驱动因素时
需要确定尽调重点以优化资源分配
安装
npx skills add mohitagw15856/pm-claude-skills --skill tornado-sensitivity -g -y
SKILL.md
Frontmatter
{
"name": "tornado-sensitivity",
"description": "Which assumption actually moves the answer — one-at-a-time sensitivity, ranked into a tornado. Use when a model's output is being argued about (LTV, ROI, forecast) and the room is debating drivers that don't matter, or before spending diligence effort: swing every driver low→high and see which one owns the outcome. Produces the ranked tornado table, share-of-swing per driver, and a real .xlsx — via the bundled zero-dependency script with a safely restricted formula evaluator."
}
Tornado Sensitivity
Every model has four drivers people argue about and one that actually controls the answer — usually not the same one. The tornado ranks them: hold everything at base, swing one driver to its low and high, measure the output range, sort. Diligence goes to the top bar; the bottom bars stop hijacking meetings.
Required Inputs
- The model — output name, a formula over named drivers (arithmetic + min/max/abs/sqrt/log/exp only), and per-driver low/base/high. The lows and highs should be defensible bounds ("the worst quarter we've seen", "the vendor's contractual ceiling"), not ±10% ritual.
- If the requester has a spreadsheet instead of a formula: extract the output cell's driver chain into a formula first, and show it for confirmation.
Output Format
- The tornado table — drivers sorted by output swing, with input range, output at each end, and share of total swing. The top driver's share is the headline ("lifetime owns 33% of the uncertainty").
- The meeting verdict — one paragraph: what deserves diligence, what deserves a decision-and-move-on, and any driver whose bounds are the real problem (huge swing because nobody actually knows the range).
- The interaction caveat — one-at-a-time ignores correlated drivers; if two move together in reality (price and churn), say so and model the pair as one driver.
Programmatic Helper
Ships scripts/tornado.py — zero dependencies, with a restricted evaluator (driver names + six math functions; anything else is rejected — injection-tested):
python3 scripts/tornado.py run tornado.xlsx --model model.json
Prints base=1.371 · top driver: lifetime (swing 1.097, 33% of total) and writes Summary + Tornado sheets. Requires a code-execution environment.
Quality Checks
- Swings computed by the script, quoted — never reasoned in prose
- Bounds provenance is stated per driver (measured / contractual / guess) — a tornado of guesses is honestly labelled one
- Share-of-swing sums are shown so the ranking's decisiveness is visible
- Correlated drivers are named and the caveat applied to them specifically
- The verdict names what to STOP arguing about — the negative guidance is half the value
Anti-Patterns
- Do not use symmetric ±X% on every driver — uniform ranges produce a tornado shaped by formula structure, not by knowledge
- Do not read the top bar as "most likely to be wrong" — it's "most consequential if wrong"; confidence and consequence are different columns
- Do not run tornado on a model whose formula the owner hasn't confirmed — sensitivity on the wrong model is confidently useless
- Do not let a huge-swing driver with made-up bounds stand — the recommendation there is "go find the real range", not "panic"
- Do not present this as risk analysis — it's attention allocation; downstream probability work still exists
版本历史
- 961cbeb 当前 2026-07-11 19:59
