tait-crt-interface-skill is a prompt template with window counts in it

Installation is a folder drop or a sentence you paste

There is no package to install, no index and no dependency to resolve. The first method is to download the entire Skill folder and place it in the local Codex skills directory at:

~/.codex/skills/tait-crt-interface-skill

The second method is to hand the job to Codex itself, by pasting a request that names the repository URL:

请帮我安装这个GitHub链接中的skill：https://github.com/TaiT-tt/tait-crt-interface-skill

Once installed, the name you invoke inside Codex is tait-crt-interface-skill, written with a dollar sign prefix in a prompt. The primary language recorded for the repository is Python, which comes from the scripts directory rather than from the skill definition itself. There are no GitHub releases, so there is nothing to pin a version to, and the declared licence is unknown, which matters if you plan to redistribute the folder.

Two questions, unless you name both answers up front

The default interaction costs two extra conversation turns. Upload a reference image and invoke the skill without naming a palette or a ratio, and the skill stops to ask. The first question is the palette, offering eight presets named 经典, 粉黛, 极客01, 极客02, 复古01, 复古02, 游戏01 and 游戏02, plus an option to reply 如图 meaning match the image. The second question is the aspect ratio, offering 3:4, 4:3, 9:16 and 16:9, and explicitly accepting any other ratio you type. Both answers can be given in the opening prompt instead, which produces an image on the first turn:

使用 $tait-crt-interface-skill 生成图片，极客02色卡，4:3比例

That single sentence is the difference between one generation and three. The presets are a closed set with one escape hatch, while the ratio is an open field.

Three style templates override a blueprint that otherwise randomises

Three representative style templates are built in, named 街头怪诞, 巨像符号 and 冷面几何. Any of the three can be written straight into a prompt to steer the output toward that style, and they compose with the palette and ratio without replacing them:

使用 $tait-crt-interface-skill 生成图片，街头怪诞，极客02色卡，4:3比例

The half worth reading is what happens when you name no template. In that case the skill falls back to its own style blueprint together with a random variant engine, so two identical prompts do not produce identical images. The template is therefore a switch rather than a requirement, and the absence of a name is a defined state rather than an error. That fallback is also the reason a reproducible result needs the template named explicitly, since the variant engine is doing the opposite of reproducible by design.

Your prompt outranks almost every setting the skill defines

The README invites extension rather than treating the built-in rules as a contract, and names the mechanism: the user instruction is set as the first hierarchy level and can override most of the built-in settings of this skill. The suggested extensions follow from that. Generate an abstract visual concept from a pure text description with no reference image at all. Produce a retro poster carrying specific text you supply. Choose the analysis type shown inside a popup window, with 解剖 for anatomy, 六角战力图 for a hexagonal power chart, 运势 for fortune, 场记板 for a call sheet and 化学式 for a chemical formula among them. The important consequence is that the specification described in the next section is a default, not an invariant, and a user who contradicts it will get their reading rather than an error.

The composition rules are counts, not adjectives

The visual specification is unusually concrete, which is the main reason to use this skill rather than describing CRT nostalgia yourself. A pixel-art subject occupies most of the frame and reads as a borderless system wallpaper. The interface language comes from early Macintosh, Minitel and 8-bit computers. The layout calls for three to six floating windows of differing size, proportion and position, plus one to three element-extraction windows reserved for facial features, accessories or local features. Subject, text, icons and windows all sit on one shared square pixel grid, and the rendering asks for checkerboard optical greyscale, hard-edged aliasing, scanlines, glow, noise and signal interference. Fixed CRT barrel distortion applies around all four sides rather than as a local effect. Colour comes from preset palettes, or the skill derives two to five colours automatically from the uploaded image.

The documented file structure omits a directory that exists

The README lists five roles. SKILL.md holds the complete workflow, the composition rules and the quality thresholds, which makes it the actual specification rather than the README. README.md is the bilingual usage guide in simplified Chinese and English. assets/ holds interactive assets such as the palette shown during the first question. references/ holds the colour and requirement specifications. scripts/ holds the shared implementation work, covering the pixel grid, palette constraints, CRT distortion and fixed signature handling. The repository root also contains an agents/ directory, which none of those five bullets mention. Nothing in the README says what lives there or when the skill uses it, so the one part of the tree you cannot reason about is the part you would expect to describe how the skill dispatches work.

A committed .DS_Store and a directory named in Chinese

Two things in the root say more about how this repository is maintained than any documented process. The first is .DS_Store, committed to version control. It is a macOS Finder metadata file that records how a folder looked when it was last opened, and it belongs in .gitignore. Its presence tells you that files were staged from a desktop Finder rather than through a filtered add, which is a small signal about how the rest of the tree was assembled. The second is the examples directory, whose name is 生成示例, meaning generated examples. A directory name in a non-ASCII script is entirely reasonable in a repository whose documentation is written in simplified Chinese, and it is also a real portability wrinkle for any tooling that assumes ASCII paths. Neither issue affects the skill, and both are worth noticing before you fork it.

Editorial conclusion

Use this skill if you want a specific retro CRT composition rather than a vague aesthetic nudge, since its value is the written specification of window counts, a shared pixel grid, checkerboard greyscale and fixed barrel distortion, all of which you would otherwise have to prompt for repeatedly. Install it by dropping the folder into your Codex skills directory, and always name the palette and ratio in your first prompt to avoid the two extra conversation rounds. Before extending it, read the reference on color and requirement specifications, and note that the repository carries no declared licence.

Frequently asked questions

Official sources

1. Issues
2. README
3. TaiT-tt/tait-crt-interface-skill on GitHub