Direct AI game screens vs controlled game screens: what changed?

劉寅 — Sun, 03 May 2026 07:48:09 +0000

Strong image models can already produce polished game UI screenshots.

The harder question is whether those screenshots are useful as production evidence.

I tested six common game-screen cases two ways:

a direct prompt baseline
a controlled workflow using a screen brief, layout contract, style contract, IP/lookalike gate, locked prompt, review score, revision prompt, and implementation notes

The six cases were:

match-3 HUD
card battle
tower defense
SLG map
roguelike reward
narrative choice

The result was not a simple "controlled is always prettier" claim.

The more useful result was narrower:

the controlled version kept gameplay state clearer
the layout was easier to review
UI hierarchy was easier to explain
rejected lookalike cases were caught before publication
implementation notes were easier to write after review

That matters because many AI game UI images fail after the image step. They look polished, but they do not tell a developer what state the screen is in, which elements are runtime text, which areas are safe for localization, or what needs to be rebuilt as components.

The workflow I am testing is:

SCREEN_BRIEF
-> LAYOUT_CONTRACT
-> VISUAL_STYLE_CONTRACT
-> IP_SIMILARITY_GATE
-> IMAGE_PROMPT_LOCKED
-> IMAGE_REVIEW
-> REVISION_PROMPT
-> IMPLEMENTATION_NOTES

The full case article is here:

https://hakurokudo.com/tools/direct-vs-skill-controlled-game-screens.html

There is also a free Sample Pack with real images, review notes, and rejected-case boundaries:

https://hakurokudo.com/tools/game-screen-generation-control.html

The paid beta is available on Gumroad and itch.io if this matches your workflow:

Boundary note: this is a workflow/control Skill, not a UI asset pack, not a copyright guarantee, and not a promise that generated images can be used directly as final game assets.

Release Guard Lite — 一个 10 秒搞定发版风险检查的 CLI 工具

劉寅 — Wed, 08 Apr 2026 15:44:02 +0000

最近做了一个小工具，解决自己发版前总是忘记检查的问题。分享给大家看看有没有用。

痛点：

用 AI 写代码（Copilot / Cursor / Claude），改了什么文件有时候自己都不清楚
小团队没有专门的 code review 流程
发版前靠记忆检查，容易漏

Release Guard Lite 做的事情很简单：

读 git diff，看改了哪些文件
检测高风险文件（数据库迁移、认证、支付、环境变量）
跑你配置的检查命令（测试、lint、类型检查）
给出风险等级：LOW / MEDIUM / HIGH
生成 Markdown 报告，可选推送到 Discord

npx release-guard-lite init   # 生成配置
npx release-guard-lite run    # 执行检查

风险判断是确定性规则，不是 AI 猜测。配置就一个 JSON 文件，零学习成本。

CI 里也能用，medium/high 时退出码为 1，可以直接阻断合并。

技术栈：Node.js 20 + TypeScript + simple-git + execa + zod，没有重依赖。

MIT 开源：https://github.com/RyuuTora1986/release-guard-lite
npm: npm install -g release-guard-lite

欢迎反馈，特别想知道大家觉得还缺什么功能。

DEV Community: 劉寅

Direct AI game screens vs controlled game screens: what changed?

Release Guard Lite — 一个 10 秒搞定发版风险检查的 CLI 工具