跪拜 Guibai
← Back to the summary

No Manual Images for 103 Mini-Games: Automating Screenshots, Covers, Packaging, and Store Submission

After my previous post "I used AI to make 103 mini-games in four days, and actually got them listed on TapTap" went out, a lot of people asked about the release details.

The truth is: writing the games took 4 days, and the release hassle took no less time than the code. But I automated all of that too.

This article breaks down the four-layer pipeline: screenshots, covers, packaging, and submission, plus the real lessons from three rejected submissions.

Results first

As of writing this article:

The essence of release is not "upload a package" but serving each platform's own asset formats. I put together a requirements matrix:

Platform Icon/Cover Screenshots/Banner Package Other
TapTap 1:1 icon At least 3 screenshots H5 zip (first <5MB) Privacy policy link, age rating, AI content declaration
GameDistribution 512×384 512×512, 200×120, 1280×720, 1280×550, 5 total Root must have index.html Manual QA
GameMonetize 512×340 minimum Multiple cover sizes Root index.html Backend iframe ad verification
HuZhan 350×280 thumbnail Detail long image Source zip + demo site Safety code secondary verification
Xianyu 750×750 first image Optional detail images Send file in chat Emoji banned in description

The same game needs 5 cover specs + 4 screenshot specs + 3 packaging formats. 103 games means a thousand images. Manual image editing is certain death, so from the first day of design, assets are part of the pipeline.

Layer 1: Automated screenshots

Gameplay screenshots use headless Chrome, one command:

chrome --headless=new --disable-gpu --hide-scrollbars \
  --force-device-scale-factor=1 \
  --window-size=512,384 \
  --screenshot=cover-512-384.png \
  "https://shuangyueliao.github.io/game-arcade/tunnel-run/"

Three pitfalls from actual testing:

  1. Minimum window width is 500px: Chrome headless has a lower limit on window-size; small sizes like --window-size=200,120 will return cropped images. The solution is to screenshot at the target size and then scale with sips/ffmpeg, don't fight the window size.
  2. Screenshot timing: Since the games render on Canvas, --virtual-time-budget=4000 lets the page run for 4 seconds before screenshotting, avoiding white loading screens.
  3. State variety: Stores want "in-game" screenshots, not the "start menu". Every game has a test hook left in the code (e.g., window.__tr exposes score and state), and the screenshot script can inject JS to jump to an obstacle- or combo-filled "good-looking" state before capturing.

Expanding on the second point: giving the game an automation hook is not a burden; it's infrastructure for the release pipeline. Screenshots, automated tests, and debugging all rely on it.

Layer 2: Cover cards are HTML, not design files

A store cover card is essentially "big text + game screen + price/selling point". Writing that in HTML is ten times faster than opening Photoshop:

<div class="stage" style="width:750px;height:750px">
  <div class="h1">Water Drop Sort<br><em>Liquid puzzle</em></div>
  <div class="feat"><span>300+ levels</span><span>Zero ad interruptions</span></div>
  <div class="p1">¥0</div>
</div>

Write one cover template; each game only changes three things: game name, one selling point, dominant color. Then headless Chrome screenshots it again: 750×750 (Xianyu), 512×384 (GD), 350×280 (HuZhan thumbnail) all come from one template, only CSS changes.

The extra benefit: cover styles are naturally uniform, so the store homepage looks like "one studio's work" rather than a grab bag of assets.

Layer 3: Packaging: one directory convention solves three platforms

game-dir/
  index.html        ← all code, single file
  (no other files)

All games are forced to be single-file: Canvas draws programmatically, WebAudio synthesizes sound effects live, zero third-party assets. The packaging freedom this constraint brings:

One zip works on every platform, and the packaging script is left only with "change store copy + swap description".

But one key detail: platform SDKs must load conditionally by domain. Each game has a snippet like this:

if (host.indexOf("gamedistribution") !== -1 || host.indexOf("gd.games") !== -1) {
  // load GD SDK
} else {
  // standalone: zero requests, zero dependencies
}

On a platform domain, the same zip automatically picks up the SDK; on GitHub Pages or locally it's the clean version. When selling source code, buyers never have to worry about "does this contain my account ID" because the condition fails and nothing loads.

Layer 4: Form submission automation

Releasing 103 games × 5 platforms also makes form filling repetitive. There's a universal solution for file uploads in web forms: DataTransfer injection.

const resp = await fetch(coverUrl);
const blob = await resp.blob();
const dt = new DataTransfer();
dt.items.add(new File([blob], "cover.png", { type: "image/png" }));
input.files = dt.files;
input.dispatchEvent(new Event("change", { bubbles: true }));

But note: not all upload components accept this. GameMonetize uses Dropzone.js, which completely ignores synthetic events. The workaround is to bypass the component and fetch POST FormData straight to its upload endpoint. The same goes for React antd/MUI dropdowns: synthetic events won't work; use coordinate-based clicks instead.

Real lessons from 3 rejected submissions

This section is the most valuable, bought with real time:

1. TapTap's first-release quota A new game's first release uses up the platform's "first-issue quota": 1 in 24 hours, 3 in 168 hours. I only found out when the second submission got blocked. Workaround: stagger the submission times of the three games.

2. Xianyu bans emoji in item descriptions Characters like ✔ and ✅ were directly rejected: "Item descriptions cannot contain emoji." I replaced them all with √ and ·.

3. Bundled package vs. single product Listing "10-game source bundle for ¥30" on HuZhan got manually rejected: "Selling multiple sets as one bundle is strictly forbidden." Listing each game as a separate product passed. On the flip side, Xianyu actually likes that bundle format. The same goods require completely opposite product forms on different platforms.

There's also a channel that didn't work out, said honestly: GameMonetize's tests judged our batch of AI-generated mini-games as "quality not meeting the bar", and with a history on the account, it refused us entirely. Conclusion: drop that channel for now, shift focus to TapTap and source sales. That's why the current focus is three TapTap games plus source-code sales.

Current form

The pipeline's source organization and release checklist have been compiled into a manual, which I'll publish once the series is done. Questions? Leave them in the comments.


Previous: 《I used AI to make 103 mini-games in four days, and actually got them listed on TapTap》