Purpose
Convert qualitative or rough probability and impact estimates for a specific risk event into a defensible expected-loss figure — a single point estimate, a P10/P50/P90 range, and a sensitivity table showing which assumptions the number is most exposed to. Replaces heat-map hand-waving with a number a CFO or risk committee can actually put in a model.
When to Use
Invoke this skill when the user asks to:
- Size a specific risk in dollar terms (expected loss, VaR-style range, or worst-case)
- Convert probability estimates and impact ranges into an expected-value calculation
- Build an expected-loss line item for a risk register or capital reserve
- Decide how much to spend on a mitigation by comparing it to the risk it removes
- Stress-test which assumption (probability, exposure, or severity) drives a risk's cost
- Compare two or three risk events on a common quantified basis
Inputs Required
Before starting, collect:
- Risk event: A specific, discrete event being quantified (e.g., "major customer default," "data breach," "supplier failure"), not a risk category
- Probability estimate: Likelihood of the event over a stated time horizon — as a percentage or a low/base/high range
- Time horizon: The period the probability applies to (annual is the default)
- Impact/severity range: Low, most-likely, and high dollar impact if the event occurs
- Exposure at risk (if applicable): The total amount exposed to loss (EAD), used when severity is expressed as a percentage of exposure
- Recovery/mitigation offset: Any recovery rate, insurance, or partial mitigation reducing net loss (default: none / 0% recovery)
- Currency and units: Reporting currency and whether figures are in dollars, thousands, or millions
If only a single probability or a single impact number is given, construct a low/base/high range around it (±40% on impact, and a stated low/high band on probability) and flag the assumption explicitly.
Instructions
Step 1: Structure the Loss Model
Decide which formulation fits the inputs and state it explicitly:
- Simple expected value: EL = Probability × Impact — use when the event is binary and impact is a single distribution
- Decomposed (actuarial) form: EL = PD × LGD × EAD — use when exposure-at-risk is provided and severity is a loss-given-default percentage
- PD = probability the event occurs in the horizon
- LGD = fraction of exposure lost if it occurs (1 − recovery rate)
- EAD = exposure at risk
- Scenario-weighted form: EL = Σ (Probability_i × Impact_i) — use when multiple discrete outcomes exist (e.g., minor / moderate / severe versions of the same event)
Show the chosen formula and label every variable with the input it maps to.
Step 2: Calculate the Point Estimate
Compute the base-case expected loss using the most-likely probability and most-likely impact (or PD × LGD × EAD). Show the arithmetic on one line so it is auditable. This is the number that goes on the register as the annual expected loss.
Step 3: Build the Probabilistic Range (P10 / P50 / P90)
Produce three loss figures reflecting uncertainty in the inputs:
- P50 (median): Base-case probability × most-likely impact — the central estimate
- P90 (adverse): High-end probability × high-end impact — a reasonable bad-case, not the absolute worst
- P10 (favorable): Low-end probability × low-end impact — the plausible good outcome Also report the conditional worst case (impact_high with the event assumed to occur, probability = 100%) separately, so the reader distinguishes "expected annual loss" from "if it actually happens."
Step 4: Sensitivity Analysis
Flex each key assumption independently while holding the others at base case, and report the change in expected loss:
- Probability sensitivity: Recompute EL at probability ±50% of base
- Severity/LGD sensitivity: Recompute EL at impact ±50% of base
- Exposure sensitivity (if EAD used): Recompute EL at EAD ±25% of base Rank the three drivers by how much each moves the expected loss. Name the single assumption the number is most sensitive to — this tells the reader where to invest in tightening the estimate.
Step 5: Interpretation and Decision Framing
Translate the numbers into a decision context:
- State the annual expected loss and the P90 in plain language
- If a mitigation cost is implied or provided, compare it to the EL reduction it would buy (mitigation is justified when its cost is below the expected loss it removes)
- Flag the largest source of estimation uncertainty and one concrete way to reduce it (better data on probability, tighter severity bounds, or confirmed recovery rate)
Output Format
- Structure: Model definition → point-estimate calculation → P10/P50/P90 range table → sensitivity table → interpretation
- Length: 500–800 words
- Format: Prose for framing, with two tables (loss range and sensitivity) and visible one-line arithmetic for each calculation
- Sections: Loss Model & Assumptions → Point Estimate → Loss Range (P10/P50/P90 + conditional worst case) → Sensitivity Analysis → Interpretation & Decision
- Tone: Quantitative risk analyst — precise, assumption-explicit, no false confidence; every number traceable to an input
Quality Checklist
Before delivering output, verify:
- [ ] The chosen formula is stated explicitly and every variable maps to a named input
- [ ] The point-estimate arithmetic is shown on one line and is reproducible from the inputs
- [ ] The range table includes P10, P50, P90, and a separately-labeled conditional worst case
- [ ] The sensitivity analysis flexes all three drivers and ranks them by impact on EL
- [ ] Every assumption made in place of a missing input is flagged explicitly
- [ ] The interpretation names the single largest driver of the number and one way to tighten it
Notes
- Expected loss (an annual average) and conditional worst case (loss if the event occurs) are different numbers for different purposes — never blend them into one figure
- When probability estimates are soft, the sensitivity analysis matters more than the point estimate; a number that swings 5x on a probability guess should be labeled low-confidence
- For rare, high-severity events (low PD, high LGD), the expected value understates decision-relevant risk — surface the P90 and worst case prominently rather than leading with EL alone
- Recovery/insurance offsets belong in LGD, not applied twice; confirm they aren't already netted out of the impact range provided
How to use this skill
This file is a Claude Code skill in Anthropic's standard format — markdown with YAML frontmatter at the top.
Installing in Claude Code:
~/.claude/skills/<folder-name>/SKILL.md— auto-loads in every session anywhere on your machine (recommended for skills you'll use across multiple projects)<your-project>/.claude/skills/<folder-name>/SKILL.md— auto-loads only when working in that project (recommended for project-specific skills)
Restart Claude Code or open a new session to register.
Using in Claude.ai, ChatGPT, Gemini, or other AI tools:
- Drag this file into a new chat, or copy and paste its contents
- Send your task or request as your next message — the AI will apply the skill where it fits
For Claude Code skill details: https://docs.anthropic.com/en/docs/claude-code/skills