How I auto-generate 800+ App Store screenshots across 39 languages and 3 devices

App Store screenshots are the highest-leverage marketing asset an app has — and the most painful to maintain. Now multiply that pain by 39 languages and 3 device classes. Doing that by hand is not "tedious," it's impossible to keep in sync.

So I built a pipeline that turns one command into ~800 finished, captioned, device-correct screenshots for Cadento, my SwiftUI focus timer. Here's the architecture.

The scale problem

The output target:

That's 819 images, each needing the right language UI and the right localized caption. Change one screen design and every number above regenerates. Hand-editing is off the table — the only sane answer is "rebuild everything from source on demand."

The pipeline, end to end

XCUITest (per language)  →  raw localized PNGs
        ↓
extract from .xcresult
        ↓
Python + Pillow: compose background + device frame + caption
        ↓
AppStore画像/<device>/<lang>/1..N.png   (exact store dimensions)

Five stages. Each is independently re-runnable.

Stage 1 — Capture real localized screens with XCUITest

The key insight: don't fake screenshots, drive the real app. A UI test launches the app, forces a specific language/locale, navigates to each screen, and snapshots it.

Language and locale come in as environment variables so one test file covers every language:

let lang   = ProcessInfo.processInfo.environment["SHOT_LANG"]   ?? "en"
let locale = ProcessInfo.processInfo.environment["SHOT_LOCALE"] ?? "en_US"

app.launchArguments += ["-AppleLanguages", "(\(lang))"]
app.launchArguments += ["-AppleLocale", locale]
app.launch()

// navigate + snapshot each screen
let shot = XCTAttachment(screenshot: app.screenshot())
shot.lifetime = .keepAlways
add(shot)

A shell loop runs this once per language. Because it's the actual app, the screenshots are guaranteed to match what users see — including RTL flips for Arabic/Hebrew and text expansion in German.

Stage 2 — Extract PNGs from the .xcresult

XCUITest buries screenshots inside an .xcresult bundle. A small Python script walks the result and pulls out the raw PNGs into a flat per-language folder. Nothing clever — just plumbing so the next stage has clean inputs.

Stage 3 — Compose with Python + Pillow

This is where raw screens become marketing. For each shot, Pillow:

1. Draws the branded background (generated separately, app-themed gradients)
2. Places the device frame
3. Drops the raw screenshot into the frame at the correct offset
4. Renders the localized caption on top — pulled from a per-language strings map

The caption text is itself localized (39 languages of ASO copy), so the marketing message reads natively, not just the UI underneath it. Font fallback matters here: CJK, Arabic, Hebrew, Thai, and Devanagari all need the right font or you get tofu (□□□).

Stage 4 — Live Activity & Watch shots

Live Activity (Dynamic Island / lock screen) and Apple Watch screens are generated through their own paths and folded into the same compositor, so the final set is consistent across all surfaces.

Stage 5 — Output to exact store dimensions

Everything lands in a predictable tree at the exact pixel sizes App Store Connect requires:

AppStore画像/iPhone_6.9/<lang>/1..10.png   (1320×2868)
AppStore画像/iPad_13/<lang>/1..8.png       (2064×2752)
AppStore画像/AppleWatch/<lang>/1..3.png    (410×502)

From here it's a straight upload (I drive App Store Connect's API to swap a single device's set without touching the others — but that's another post).

Lessons from running it for real

• Drive the real app, don't mock. The whole value is that screenshots can't lie about what the UI does in each language.
• Environment variables > 39 test targets. One parameterized UI test beats copy-pasted code every time.
• Font fallback is not optional. Test the hardest scripts (Arabic, Thai, Hindi, CJK) early or you'll ship boxes.
• Make every stage idempotent. A design change should be one command away from 819 fresh images, not a weekend.
• Separate UI capture from caption rendering. Redesign the screen? Re-run stage 1. Rewrite the marketing copy? Re-run stage 3. They shouldn't be coupled.

The payoff: when I change a screen or a tagline, I'm not dreading a manual marathon. I run the pipeline, and the entire localized store presence updates itself.

I'm a solo iOS developer from Japan building small, deeply localized apps. Cadento (focus timer, 39 languages) is on the App Store. Ask me anything about the pipeline in the comments.