--- name: promotion-campaign-review description: Design or audit one promotion system—offer, discount, update benefit, cross-promotion, win-back, lifecycle push, app/game event, referral push, launch or seasonal campaign—across objective, audience, deterministic eligibility, user+offer cooldowns, placement, transparent message, economics, authoritative fulfillment, reversal, fraud, support, experiments, automation, and shutdown. Use for one campaign artifact; use Marketing Automation for the whole channel and spend operating system. --- # Promotion Campaign Review Produce a **Promotion Campaign Contract** that grows incremental retained value without fake urgency, loyal-customer punishment, review manipulation, spam, or unreconciled rewards. ## Atomic boundary Own one campaign/offer's objective, audience, eligibility, placement, message, economics, applicable benefit fulfillment/reversal, channel handoffs, fraud, support, experiment, and shutdown. Do not own the whole marketing system, payment provider ledger, referral program, store listing, or game/app design. Use a draft artifact ID. Consume payment, refund, economy, referral, notification, analytics, marketing, and product decisions by owner and explicit contract. Let deterministic delivery tooling seal versions/digests later; never invent them during design. ## Agent-first invariant Construct every applicable campaign state, eligibility policy, channel asset, provider authority, experiment, fraud/support path, telemetry, cap, expiry, kill switch, and reconciliation before activation. Build ledger/grant/reversal adapters only for a discount, reward, credit, entitlement, or other fulfilled benefit; informational/lifecycle campaigns record that mode as non-applicable. Low volume or unknown ROI never justifies manual fulfillment or later hardening. Separate construction, exposure, qualification, spend, and any grant authority; dormant campaigns perform zero sends, SDK work, grants, or data collection. ## Workflow 1. Define exact objective and incremental-value hypothesis, audience/cohorts, platforms/territories/age modes, baseline, budget/reward/margin caps, time horizon, owner, and ruin boundaries. 2. Read `references/promotion-campaign-patterns.md`. Select campaign/benefit mode (`informational`, `discount`, `reward`, `credit`, `entitlement`, or another explicit type), state the normal alternative, and mark fulfillment/reversal branches applicable or not applicable. Model loyal/active, new, lapsed, payer/non-payer, ownership, refund/dispute, consent, region, platform, and abuse states. 3. Define deterministic view and exposure eligibility, plus redeem/qualify/grant eligibility only when the selected benefit requires it, with reason codes, user+offer cooldowns, conflict/stacking, caps, expiry, ownership overlap, baseline path, and support explanation. 4. Place only at a value/intent transition. Specify cross-channel dedupe, dismissal/suppression, quiet hours, preference/consent, privacy/redaction, deep link, and notification/store/referral handoffs. 5. Write transparent benefit, condition, duration, renewal/normal state, limits, expiry, reversal, and support route. Localize meaning, not just strings. 6. Model exposure and, when applicable, authoritative fulfillment separately, including no-op, failure, duplicate/out-of-order, refund/chargeback, fraud, rollback, compensation, restore, migration, support correction, and live readback. 7. Compute incremental retained contribution after discount/reward, fees, cannibalization, refunds/chargebacks, fraud, fatigue, support, economy, and loyal-customer effects. Define randomized/causal cells where feasible. 8. Set exact scale/hold/pause/withdraw predicates, provider-side end times, upstream caps, emergency shutdown, reconciliation-only withdrawal state, and post-campaign closeout. ## Source verification Retrieve current store/payment/promotion, advertising, notification, consent, child/age, referral, review/rating, pricing, tax, territory, and channel policy for the exact campaign. Explicit prohibitions and missing authority are hard floors. Genuinely ambiguous but admissible behavior may use the bounded risk contract below; known policy violations, deception, consent bypass, child harm, or unlawful behavior are not “grey experiments.” ## Bounded risk-reward contract Use arithmetic to compare admissible options without pretending fat-tail or irreversible harm is an ordinary average cost: ```text expected_net_value = P(success) * incremental_retained_contribution - sum(P(loss_i) * direct_and_secondary_loss_i) - trust_and_reputation_cost - recovery_and_opportunity_cost constraints: maximum_loss <= declared_ruin_budget CVaR_at_selected_confidence <= declared_risk_budget exposure <= blast_radius_cap time_to_detect + time_to_contain <= recovery_window ``` Record probability ranges and sensitivity rather than false precision. Include platform/account action, customer remedy, refunds, support, reputation, cannibalization, reversibility, detection lag, rollback limits, and opportunity cost. Unknown material facts widen the loss range; they do not become zero. An experiment may proceed only when the behavior is not explicitly prohibited, the customer proposition is truthful, consent and remedies remain intact, exposure/spend are capped, stop signals are observable, and recovery has been tested. A small company’s ability to correct quickly lowers some recovery cost; it does not erase platform-account, legal, child-safety, privacy, or trust ruin. ## When not to use - Use `marketing-automation-blueprint` for the full organic/lifecycle/paid, creative, attribution, spend, reputation, and shutdown operating system. - Use `referral-loop-review` for persistent inviter/invitee qualification, attribution, pending grants, reversals, fraud, and support. - Use `store-listing-optimization` when the artifact is one channel's metadata, screenshot/video/capsule sequence, proof, localization, and conversion test. - Use `daily-reward-and-streak-review` for a durable recurring return-loop state machine, or `game-economy-review` for economy balance and inflation. ## Guardrails - No fake scarcity/countdown, hidden renewal/normal price, shame, blocked exit, reset timer, or loss threat involving creations/history/identity/access. - Never reward review/rating, permission, ad click, spend, forced referral, or mere update installation. Update benefit requires verified new-value use. - A user cannot see an offer they cannot redeem; regional/platform differences and suppression are support-explainable. - Never confirm reward before authority and idempotent grant commit. Ordinary refund/reversal cannot delete unrelated value or auto-ban an account. - Optimize neither gross revenue nor clicks alone. Retention, renewal, margin, refunds, churn, loyal cohort, fatigue, support, fairness, economy, privacy, and fraud false positives are mandatory countermetrics. - Autonomous optimizers cannot alter eligibility floors, prices/terms, consent, reward value, grant/reversal, spend caps, or their own promotion gates. ## Output contract Return one typed Promotion Campaign Contract with: 1. objective, audience/cohort, baseline, exact offer, economics, authority, budgets/caps, horizon, assumptions, and ruin boundaries; 2. view/exposure and applicable redeem/qualify/grant eligibility, reason, conflict/stacking, user+offer cooldown, ownership, expiry, and support matrix; 3. placement/message/localization, channel dedupe/suppression, consent, preference, privacy, and deep-link contract; 4. exposure state machine plus applicable fulfillment/reversal/compensation state machines and ledger, or an explicit non-benefit disposition; 5. applicable fraud, refund/dispute, restore/migration, support, and reconciliation rules; 6. event schema, causal experiment, economics and trust countermetrics; 7. expected-value/CVaR record where uncertainty is material, plus scale/hold/pause/withdraw, provider end/cap, kill switch, live readback, and closeout evidence; 8. specialist handoffs with draft IDs, owners, required inputs/outputs, acceptance questions, and no fabricated proof. Complete only when every eligible/ineligible impression, applicable qualification/grant/reversal, and withdrawal is explainable, bounded, replayable where applicable, and support-safe.