Purpose
Produce a defensible credit risk segmentation framework that divides a lending population into distinct, treatable segments — each defined by concrete bureau, income, and behavioral cutoffs and paired with a specific underwriting treatment, pricing tier, and credit line strategy. Replaces single-threshold or score-only decisioning with differentiated policy that improves approval rates in low-risk pockets while tightening exposure where loss rates concentrate.
When to Use
Invoke this skill when the user asks to:
- Build or redesign a credit segmentation scheme for an underwriting policy
- Define risk-based pricing or credit line assignment tiers
- Differentiate underwriting treatment across thin-file, prime, near-prime, and subprime populations
- Translate a credit score and bureau attributes into actionable approve/decline/refer logic
- Refresh segmentation after a portfolio shift, new product launch, or rising delinquency
- Justify a segmentation strategy to a credit committee or model risk function
Inputs Required
Before starting, collect:
- Product type: Card / personal loan / BNPL / line of credit / auto, including typical term and amount range
- Population: New applicants, existing portfolio, or both; approximate volume and geography
- Available data: Which bureau attributes (FICO/VantageScore, trade lines, utilization, inquiries, delinquency history), income/affordability signals (stated, verified, cash-flow), and behavioral data (prior account performance, app activity) are accessible
- Risk appetite: Target loss rate, target approval rate, or both
- Current decisioning: Existing cutoffs or policy in place, if any
- Regulatory constraints: ECOA/fair lending considerations, adverse action requirements, jurisdiction
- Economics: Approximate revenue per good account and loss per bad account, to weigh segment trade-offs
If loss/approval targets or unit economics are missing, default to industry-typical bands for the stated product and flag the assumption explicitly.
Instructions
Step 1: Define the Segmentation Dimensions
Select 3 primary axes that will structure the framework. Do not exceed three primary axes — interaction complexity grows multiplicatively.
- Risk tier (bureau-driven): Bucket the credit score into bands aligned to the population (e.g., Super-prime 760+, Prime 700–759, Near-prime 640–699, Subprime 580–639, Deep subprime <580). Adjust band edges to the product and risk appetite.
- Affordability/capacity (income-driven): Use debt-to-income, payment-to-income, or cash-flow signals where available; where income is unverified, note reliance on bureau-derived obligation estimates.
- Behavioral/stability: File depth (thin vs. thick), recent inquiry velocity, delinquency recency, utilization trend, or prior account performance for existing customers.
State which dimension is primary (drives the coarsest split) and which are secondary modifiers.
Step 2: Construct 6-10 Segments With Explicit Cutoffs
Build the segment grid from the dimensions in Step 1. For each segment, specify:
- Segment name and ID (e.g., S3 — "Near-prime, thin-file, low utilization")
- Defining rule: Concrete cutoffs in plain logic (e.g., "score 640–699 AND DTI ≤ 35% AND ≤2 inquiries in 6 months AND no 60+ DPD in 24 months")
- Expected population share: Approximate % of volume falling into the segment
- Risk hypothesis: Why this group behaves distinctly — the loss driver or protective factor at work
Ensure segments are mutually exclusive and collectively exhaustive (every applicant lands in exactly one). Include an explicit catch-all / referral segment for cases that fail clean classification.
Step 3: Assign a Differentiated Treatment Per Segment
For each segment, define the decisioning treatment:
- Decision: Auto-approve / approve with conditions / manual referral / auto-decline
- Pricing tier: APR band or risk premium relative to base
- Credit line / loan amount: Assignment rule (fixed, multiple-of-income, score-graduated)
- Conditions or step-ups: Income verification trigger, deposit/collateral requirement, lower starting line with graduation path
- Adverse action mapping: For decline/conditional segments, the principal reason codes that apply
Step 4: Stress-Test the Framework
Validate the design before finalizing:
- Edge cases: Walk 3 borderline applicant profiles through the grid and confirm each lands in a sensible segment.
- Cliff risk: Identify cutoffs where a 1-point score change flips approve→decline; flag whether a buffer or referral band is warranted.
- Fair lending check: Confirm no segment rule uses a prohibited basis or an obvious proxy; flag any attribute (e.g., zip-code-derived signals) requiring disparate-impact review.
- Approval/loss reconciliation: Estimate blended approval rate and loss rate across segments against the stated risk appetite; note the gap if targets are not met.
Step 5: Define Monitoring and Refresh Triggers
Specify how the segmentation is governed post-launch:
- Segment-level KPIs: Approval rate, early delinquency (30+ DPD at 3/6 months), vintage loss curve, population drift per segment
- Drift thresholds: Population-share shift that should trigger review
- Refresh cadence: When to recalibrate cutoffs (scheduled vs. event-driven)
Output Format
- Structure: Narrative framing of dimensions, then a segment table as the core deliverable, followed by treatment matrix, stress-test findings, and monitoring plan
- Length: 1,000–1,400 words
- Format: Mix of a structured segment/treatment table and supporting prose; cutoffs expressed as explicit logic rules
- Sections: Segmentation dimensions → Segment definitions table (6–10 rows) → Treatment matrix → Stress-test findings → Monitoring & refresh triggers → Assumptions and caveats
- Tone: Credit risk specialist — quantitative, policy-precise, explicit about trade-offs and regulatory exposure
Quality Checklist
Before delivering output, verify:
- [ ] Between 6 and 10 segments are defined (count them), including a catch-all/referral segment
- [ ] Segments are mutually exclusive and collectively exhaustive — no applicant fits two segments or none
- [ ] Every segment has concrete numeric cutoffs, not qualitative labels alone
- [ ] Every segment maps to a specific decision, pricing tier, and line/amount rule
- [ ] A fair lending check is performed and any proxy-attribute risk is flagged
- [ ] Blended approval and loss estimates are reconciled against the stated risk appetite
- [ ] Decline and conditional segments include adverse action reason-code mapping
Notes
- Score-only segmentation underperforms because two applicants at the same score can carry very different affordability and behavioral risk — the value of this skill is in the second and third axes.
- Resist over-segmentation: more than ~10 segments fragments volume, weakens statistical reliability of monitoring, and complicates governance without improving discrimination.
- Where income is stated rather than verified, treat affordability cutoffs as softer and lean on verification step-ups in the treatment layer rather than hard declines.
- Always recommend validating proposed cutoffs against historical vintage performance before production deployment — this framework is a policy design, not a substitute for backtesting.
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