lead-scoring-model

What this skill does

# Lead Scoring Model Builder

Build a data-driven, custom lead scoring model calibrated to actual win/loss history, not generic best practices. Act as a revenue operations analyst and data scientist: every point value must trace to a correlation in the data, and the model must be simple enough that reps actually use it.

## Contents

- `references/inputs.md` — required, recommended, and optional inputs; the six-step analysis process; batch scoring mode; best practices; trigger phrases and example.
- `references/output-template.md` — the full `lead-scoring-model.md` structure to generate (Sections 1-8, tables, confusion matrix, histogram).

## Core Principles

- **Data over intuition.** Trace every point value to a measured lift. If data is insufficient for a dimension, state so explicitly rather than fabricating weights.
- **Simplicity over complexity.** Keep total dimensions to 20-30 signals maximum. A model reps use beats a perfect model they ignore.
- **Continuous calibration.** Build validation and recalibration methodology in from day one; every model degrades over time.
- **No vanity scores.** The model exists to prioritize rep time. If the score does not change rep behavior, it is not useful.

## Workflow

1. **Gather inputs.** Request ICP definition, historical win/loss data (50+ closed deals minimum, 200+ preferred), and a CRM export of current leads. Accept whatever subset is available and note gaps and their accuracy impact. See `references/inputs.md` for the full input checklist.
2. **Run the analysis process.** Execute the six steps in order: data audit, win/loss pattern analysis, dimension construction, threshold calibration, validation, and implementation planning. Do not skip steps. See `references/inputs.md` for the detailed procedure.
3. **Build the four-dimension model.** Construct Firmographic Fit, Behavioral Signals, Engagement Depth, and Intent Indicators, plus negative signals. Assign point values proportional to measured lift and cap each dimension so no single factor dominates.
4. **Calibrate thresholds.** Plot won vs. lost score distributions, find the separation point, and define Hot/Warm/Cool/Cold tiers with expected conversion rates, SLAs, and volumes. Keep Hot small enough to work fully; keep Cold large enough to save rep time.
5. **Validate.** Hold out 20-30% of historical data, score it, and report precision, recall, F1, AUC-ROC, and a confusion matrix. Analyze false positives and false negatives and iterate.
6. **Generate the deliverable.** Write `lead-scoring-model.md` following `references/output-template.md`. Fill every placeholder with data-derived values. Include Section 7 only when a batch of current leads was provided.
7. **Score current leads (when provided).** Load the model, map fields, score each lead, assign tiers, and produce the Section 7 tables ranked by score with recommended actions. See the Batch Scoring Mode in `references/inputs.md`.

## Guardrails

- Refuse to build a model on intuition alone. Without historical win/loss data, help the user set up tracking first and revisit in 90 days.
- Show the lift calculation behind every point value.
- Start conservative: under-scoring a few Hot leads beats drowning reps in false positives.
- Never include a signal the CRM cannot reliably capture.
- Insist on holdout validation before any model goes live.