Build a Year-in-Review video generator (Spotify-Wrapped style)

Every December for the last five years, Spotify Wrapped has eaten a week of cultural attention. The format has now generalized — Strava, GitHub, Apple Music, Duolingo all ship their own per-user year-end recap. The mechanics are not magic: a small set of HTML scene templates, a CSV of per-user stats, a deterministic renderer, and a queue.

Here is how it actually fits together.

The architecture

Three pieces:

1. A scene library. 6-10 HTML scenes, each parameterized by user stats. "Top artist of the year." "Minutes listened." "Top genre." "Compared to last year." Each scene is a 4-6 second render(t) animation.
2. A per-user data pipeline. SQL or warehouse query that, for every user, produces a JSON blob: { topArtist, minutes, topGenre, vsLastYear, ... }.
3. A renderer that joins the two. For each user, fill the templates with their data, render each scene, concatenate into one MP4.

The fundamental insight: the scenes are the same; the data is different. Spotify's Wrapped looks personalized because it is personalized — but the underlying templates are the same for every user.

The per-scene template

A KPI-card-pop scene, parameterized:

The template knobs:

• TITLE — "Minutes listened in 2026"
• VALUE — the user's specific number
• DELTA — comparison to last year
• LABEL — small caps label
• ACCENT — brand color

The data joins at render time. The renderer pulls one row of CSV/JSON per user, substitutes the variables, and renders.

The concatenation step

Six scenes × 5 seconds = 30 seconds of video. Either:

1. Render each scene to MP4, then concatenate with ffmpeg concat. Simple, fault-tolerant, scales out naturally — each scene is an independent render job.
2. Render the whole 30-second timeline as one render(t). More efficient (one Chromium load per user), but the template is more complex.

For a first version, go with option 1. The cost is small (one extra ffmpeg invocation per user) and the debug story is much better — if scene 4 has a glitch, you re-render scene 4 only.

The scale problem

Spotify has 600 million users. You probably don't. But "year-end recap for our entire user base" is still a large enough number that the rendering has to be batched.

For 100,000 users at 30 seconds each, plan for:

• ~1.5 seconds of render per user-second of video at 1080p on a modern Chromium. So 45 render-seconds per user.
• 100,000 × 45s = 1.25 million seconds = 350 hours = 14.6 days on one machine.
• Parallelize across 100 worker machines: 3.5 hours.

The whole pipeline is described in render 10k variants overnight — same shape, just more users.

The variable choices that matter

The thing that turns a year-in-review video from "data dump" to "shareable" is the editorial pacing:

1. Start small. Open on a single quiet stat. Not the biggest number.
2. Build to a hero number. The fourth or fifth scene should be the one people screenshot.
3. End on a tagline. "Your 2026, in 30 seconds" or similar. Brand mark.
4. One color per year. Wrapped uses a different palette every year. Do the same — it makes the share-back instantly recognizable as "this year's Wrapped" vs last year's.

Editorial discipline matters more than render speed.

The CSV / JSON shape

For each user, one row:

{
  "user_id": "abc123",
  "name": "Jordan",
  "minutes_listened": 42384,
  "top_artist": "The National",
  "top_genre": "indie folk",
  "vs_last_year_pct": 18.4,
  "top_song": "Bloodbuzz Ohio"
}

The CSV is the contract between the data team and the render team. Add a column, add a scene. Remove a column, remove a scene. No coordinator meetings required.

The same pattern in plain CSV form is in batch personalized videos from CSV. The KPI scene specifically is detailed in animated KPI cards that look like money.

The headline

A year-in-review video generator is a CSV, a scene library, and a render queue. None of the three pieces are exotic. The hard part is the editorial taste of the scenes — and that part doesn't get easier with a bigger render farm. Build the scenes first; the renderer is small infrastructure work behind them.