盛永裕介

Posted on May 2 • Originally published at hogwartzinc.itch.io

I shipped a polished hanafuda card game in 2 days with Claude and Godot

#godot #gamedev #csharp #ai

I just released Shin KoiKoi — a free, polished take on the traditional Japanese card game Hanafuda Koi-Koi. Solo dev, 2 days from kickoff to itch.io public page, with full multi-language support and accessibility features.

This post is a release log. I'll walk through the technical choices, what went right, and the three or four things that bit me. No hype — just what actually happened.

Why this game

The market for hanafuda apps is bleak. Most options are ad-laden, freemium-gated, or visually frozen in 2010. I wanted a polished, fully-offline, ad-free version that respects the player. So I built one.

The scope decision that made this feasible: only the Koi-Koi rule set. No Hachi-Hachi, no Hanaawase variants, no Mushi. Single-player only for v0.1.0. Online multiplayer is planned but deferred.

The stack

Area	Choice
Engine	Godot 4.6.2 .NET (mono)
Language	C# (`net9.0` target)
AI pair programming	Anthropic Claude
Visual assets	Gemini nanobanana2 / Midjourney V7
Fonts	Noto Sans CJK + Noto Sans Devanagari (OFL)
Tests	184 unit tests (homemade xUnit-like, all green)
Distribution	itch.io Direct (Mac/Win/Linux simultaneous)

I picked Godot over Unity for three reasons: open source + zero royalties at scale, smaller binary footprint than Unity, and faster cold-edit cycle (the script-error-restart loop is much friendlier than Unity's domain reload).

What "2 days" actually means

I want to be honest about the timeline:

Pre-work: ~3 weeks of design notes, scope decisions, and rule research before I touched the editor
Active dev: 2 focused days of code + asset curation
Polish day: itch.io setup, screenshots, multilingual descriptions, SNS announcements

So "2 days" is the interesting part — actual implementation. The pre-work matters; jumping into Godot without a scope decision would have stretched this into weeks.

The 2 days included Claude writing roughly 70% of the UI scaffolding under my direction. I supplied game logic, iterations, and the "feel" decisions (animation timing, spacing, color theory). The collaboration was much closer to "principal engineer with junior pair" than "AI did it all" — Claude needed direction at every architectural decision, but the typing and boilerplate were near-instant.

Things that bit me

1. macOS export ETC2/ASTC requirement

Universal binary export for macOS fails with a non-obvious error if you don't explicitly enable VRAM compression in project settings:

ETC2 ASTC テクスチャフォーマットが無効になっている場合、
ユニバーサルまたは arm64 用にエクスポートできません

The fix is one line in project.godot:

[rendering]
textures/vram_compression/import_etc2_astc=true
textures/vram_compression/import_s3tc_bptc=true

Cost me 30 minutes searching the wrong way. Documenting it here so the next person Googles directly to the answer.

2. Multi-language font fallback (especially Hindi)

Adding Devanagari (हिन्दी) to a font stack that already covers CJK and Latin requires careful priority ordering:

var sysFont = new SystemFont
{
    FontNames = new[]
    {
        "Noto Sans CJK JP",
        "Hiragino Sans",
        "Yu Gothic",
        "Noto Sans",
        "Noto Sans Devanagari",  // ← critical for Hindi
        "sans-serif",
    },
    AllowSystemFallback = true,
};

If you put Noto Sans Devanagari before Noto Sans CJK JP, CJK characters render with weird stroke widths. Tested all 6 languages (ja/en/de/fr/es/hi) — no missing glyphs ("tofu" boxes) anywhere.

3. AI-generated screenshot automation

Manual screenshots are tedious and inconsistent. I wrote a Node-based state machine in C# that:

Initializes the Main scene
Walks through Title → Game → Yaku list → Stats → Settings
Captures viewport texture as PNG at each stop
Quits cleanly

public partial class ScreenshotTour : Node
{
    public override void _Process(double delta)
    {
        _stepElapsed += (float)delta;
        switch (_step)
        {
            case 0: if (_stepElapsed >= 3.0f) Advance(); break;
            case 1: Capture("01_title.png"); Advance(); break;
            case 2: _main.OpenSettingsForScreenshot(); Advance(); break;
            // ...
        }
    }
}

Triggered via CLI: --screenshots /tmp/output_dir. Now I can re-shoot all marketing screenshots in 15 seconds whenever the UI changes. Worth the 30 minutes to write.

4. Oracle Cloud Always Free region trap

Tried to use Oracle Cloud's Always Free tier for online multiplayer (Phase 2). Hit two roadblocks:

ARM Ampere A1.Flex shape is chronically out of capacity in Tokyo (50+ retry failures)
Subscribing to other regions (Phoenix, Osaka) requires upgrading to Pay-As-You-Go — the home region is the only one available on the strict Always Free tier

Workaround for now: ship offline-only v0.1.0 to itch.io, defer online multiplayer to v0.2 when capacity opens up or I switch hosting strategies. This was the right call — the game is complete as a single-player experience and shipping early matters more than adding multiplayer to an unproven idea.

AI art with zero IP risk

I generated ~45 PNG assets using Gemini nanobanana2 and Midjourney V7. All disclosed openly via itch.io's "AI generation disclosure" toggle (it's a yes/no field; I selected yes).

The risk-mitigation choice that made this clean: no characters, no people. Only:

Family crests (kamon)
Hanafuda card motifs (12 months of seasonal nature imagery — public-domain cultural symbols)
Backgrounds (pine, cherry blossom, moon)
Rank badges (kanji on lacquered medallions)

Every element is either a public-domain cultural symbol or an abstract decoration. No specific character likeness anywhere, so no IP collision risk.

Accessibility from day one

This is non-negotiable for me. v0.1.0 ships with:

Colorblind mode (deutan/protan-aware saturation tweaks)
4 UI scales (0.85x / 1.0x / 1.15x / 1.30x)
Reduced motion mode (shortens cutin animations)
6 languages with proper script support
GDPR/CCPA consent flow on first run

Adding these later costs more than building them in upfront, and "indie casual" isn't an excuse to skip them.

What I'd do differently

Set up automation earlier: I should have written the screenshot tour and SDK install scripts in hour 1, not hour 30
Pre-build the icon set: Apple/Google demand many icon sizes; using sips to batch-resize would have saved 20 minutes
Test on Linux earlier: I built Linux export last and discovered a permissions issue. Should have done a quick smoke test on each platform every few hours

Roadmap

✅ v0.1.0 (2026-05-02): Single-player, all 3 desktop platforms, 6 languages, accessibility
🔜 v0.1.1: Bug fixes from itch.io community feedback
🔜 v0.2: Online multiplayer (Nakama + free server hosting once capacity opens)
🔜 Mobile: Google Play ($25 one-time) and possibly App Store ($99/year if interest justifies)
🔜 Steam: Landscape UI implementation first, then Steamworks Direct ($100 per submission)

Try it

Download Shin KoiKoi (free) — Mac, Windows, Linux. No ads, no IAP, no tracking.

Source isn't public yet, but if any specific subsystem (the AI evaluator, the locale system, the screenshot tour) interests folks I'm happy to write follow-up posts. Reply or DM with what you'd want.

Built solo by hogwartz.inc. Engine: Godot 4.6.2. Pair programming: Anthropic Claude. Made with respect for the traditional game and for the player's time.

DEV Community