Spotify Wrapped is a growth loop, not a year-end gimmick. Ship one for your AI app in an afternoon.

The implementation

Spotify Wrapped works because of one psychological fact: people love seeing their own behavior reflected back as a story. It drives shares, re-engagement, and (quietly) Premium upgrades. Every AI app with usage data can run this loop monthly - "you generated 214 images this month, here's your story" - and almost none do, because turning usage rows into narrative used to be a content problem. It's now a schema problem: a monthly usage recap, generated per user, in one call. My app meters usage through Vevee (@vevee/sdk), so every user's month already exists as structured events. compose() reads that and returns the recap as typed JSON. The compose type's data sources are the interesting part: alongside the user's own events and usage, it can pull population-level blocks - popular features and aggregate comparisons - so the recap can say "you used background removal more than most users this month" without me building any cohort analytics. The intent prompt: "Write a celebratory monthly recap from this user's real numbers. Concrete over generic - counts, streaks, firsts. End with exactly one suggestion based on what power users do that this user hasn't tried."

import { createClient } from "@vevee/sdk";

const vevee = createClient({ apiKey: process.env.VEVEE_SECRET_KEY! });

interface MonthlyRecap {
  headline: string;        // "Your February in 214 images"
  highlights: string[];    // 3 concrete, number-driven facts
  topFeature: string;
  percentileBlurb: string; // "more than most users this month"
  nudge: string;           // ONE forward-looking suggestion
}

const recap = await vevee.compose<MonthlyRecap>("monthly-recap", userId);

if (recap.status === "generated") {
  renderRecapCard(recap.output);   // in-app card + email
  await vevee.capture({ distinctId: userId, event: "feature_used",
    properties: { feature: "monthly_recap" } });
}
// opted_out → no recap; respect it silently

What a user gets

Here is the rendered recap for one user's month. Every number is real, pulled from the user's metering events. The model's job is narration, not invention - the schema and data blocks keep it honest.

• Your February in 214 images
• 214 generations - your biggest month yet (January was 90)
• Background removal 61 times: your signature move
• First HD export on Feb 12 - then 23 more
• You used the editor more than most users this month. One thing you haven't touched: batch processing - power users pair it with background removal to do in minutes what you're doing in sessions.

Where the conversion hides

The recap is an engagement feature on the surface and a paywall setup underneath. Measure it with the reserved-event funnel you already have: feature_used (recap viewed) → paywall_shown → checkout_completed, filtered by properties.feature = "monthly_recap". Three mechanisms do the work:

• It quantifies the habit. A user who reads "214 generations, your biggest month yet" has just been told, by their own data, that they are a power user. That identity does more pre-selling than any feature list.
• The nudge is plan-aware. For free users near their cap, the forward-looking suggestion naturally becomes "at this pace you'll hit your limit by the 9th" - and your limit-reached upsell flow takes it from there.
• It re-engages on a schedule. A monthly email people actually open (it's about them) is a recurring re-activation surface that costs fractions of a cent per user - usage.costMicroUsd on each result keeps the receipts, and the dashboard's monthly compose budget hard-caps the spend.

Afternoon checklist

Four pieces, none of them novel. The data's been sitting in your metering layer the whole time. Wrapped was never a data problem - it was a "who's going to write 10,000 personalized summaries" problem. That problem is gone.

• One compose type in the dashboard (prompt + data sources + JSON schema) - 20 minutes
• One server route calling compose<MonthlyRecap>() - it's sk_*-key, server-side
• One recap card component + one email template
• One monthly cron iterating active users