---
name: referrals
description: Designs customer referral and affiliate incentives with qualifying events, contribution economics, and abuse controls. Use when the founder asks for refer-a-friend rewards, affiliate commissions, ambassadors, or referral incentives.
license: MIT
disable-model-invocation: true
---

# Referrals

Reward the customer outcome you can afford. Count invitations separately from acquired customers.

Run `node scripts/run-state-check.js referrals`. Requires Product truth, ICP hypothesis, Diagnosis, and Channel bet. Inspect Last observation for repeat value and unprompted recommendations. If retention is unknown, propose a small manual introduction test; do not design a large reward system on assumed loyalty.

Read [references/playbook.md](references/playbook.md). For a creator's first collaboration rather than an incentive program, recommend `partnerships` as the next skill without opening it.

## Process

1. Choose customer referral or affiliate; state who recommends, to whom, and why their reputation benefits.
2. Place the ask after an observed value event. Name the friend's first-value action and the event that qualifies a reward.
3. Draft the incentive and terms: sides, amount, caps, expiration, qualification, refunds/reversals, attribution window and conflict rule, support/disputes, and payout owner.
4. Model rewards at incremental delivery cost, including both sides, commissions, support, fraud allowance, and discounts. Separate attributed purchases from incremental ones.
5. Pilot with a small named eligible cohort before software. Include a holdout or comparable earlier cohort where feasible, and acknowledge selection bias.
6. Draft one referral invitation or affiliate offer; run `write`. Review the economics and terms with `review`. Do not recruit, pay, or send.

Write **Success / kill**; use **Last observation** only for founder-reported cohort results.

## Output

Program brief, contribution worksheet, event/attribution table, one paste-ready invitation, exact placement, and watch/stop criteria. Any numerical illustration must be labeled hypothetical.

## Done

Reward liability is capped, qualification is testable, referral CAC uses the right denominator, and no revenue projection is presented as a result.