nickture-interface

Nickture Skills for interface polish, Russian copy and AI deslop

Two Agent Skills that help an AI agent make an interface clear, consistent, easy to use and scalable, and Russian text clear and correct, without the signs of AI-generated work. Each skill is a set of checkable rules. The agent follows them when it edits an interface or text, and reviews finished work against them. They work in Claude, Codex, Cursor and other compatible agents.

The skills themselves are in Russian for now. The agent reads them as they are and replies in your language, so nickture-interface works for an interface in any language. nickture-text-ru covers Russian text only, and an English version is planned.

I’m Nick, a product and interface designer with almost thirty years of experience. I consult teams on product, interface and conversion, record video reviews of products, teach managers to tell a good interface from a bad one, and build products with my team from idea to launch. If you need any of this, contact me.

With nickture-interface the agent also does its best to remove the signs of generated design, collected in its Slop (Слоп) section. Among them are three columns with an icon in a colored circle, a blue-to-purple gradient, everything centered, a card around every block, a badge that repeats the heading next to it, monospace type where it doesn’t belong, numbering like “01 / 02 / 03” or “0.1 / 0.2 / 0.3”, even nested several levels deep, and emoji in place of icons.

With nickture-text-ru the agent edits text so it reads easily and is correct, and does its best to remove the signs of generated text. Its Slop (Слоп) section lists several dozen of them, including clichés, “not X but Y” constructions, bureaucratic phrasing, intensifiers and hedge after hedge.

The rules come from my own reviews and thinking, and from open catalogs, books and articles. The key at the end of a rule points to the source table in sources.md. In nickture-interface a star ★ marks the rules about the most visible mistakes. They are listed in the quick pass (Экспресс-проход) section, and when time is short they are checked first. A rule can be broken when there is a reason for it.

For AI agents

If you were given a link to this repository, clone it into a temporary folder outside the project and read the files from disk.

git clone --depth 1 https://github.com/nickture/skills /tmp/nickture-skills

A web fetch tool returns a short summary instead of the file, so the rules never reach you. If you can’t clone, read each file in full from raw.githubusercontent.com with a tool that returns it unchanged. If that is impossible too, tell the user that the rules were not read in full.

To review a finished project, read SKILL.md of each skill and every file in its rules/ folder in full. Review the text and the layout in separate runs, each in its own agent or session. Check the text against nickture-text-ru first, then the layout against nickture-interface on the rendered page. For an edit, read SKILL.md, the slop sections and the sections the edit touches, and check your result against them. Take the product from an exact copy: its files, or the live page opened in a browser. Download the page with curl for the search scripts, or instead of the browser only when none of your tools can open one. A summary of the page is not enough.

Installation

Install the skills or clone the repository. A link to it in a prompt is not enough, because the agent reads the link through a web fetch tool, which passes on a summary instead of the rules.

From the Claude Directory

The plugin is in the Claude Directory. Open the plugin page on claude.ai and select Add. The plugin then works in the Claude app and in Cowork. It also loads in Claude Code when you sign in there with the same account, in version 2.1.273 or later. This is the only way the plugin updates by itself without any setup. A new version reaches the directory after Anthropic reviews it, so the directory can be a few versions behind GitHub.

To install from the directory inside Claude Code, run this command:

/plugin install nickture-skills@anthropic-plugin-directory

Claude Code has the directory built in, so you don’t add a marketplace first.

With the skills CLI

It installs skills into Claude Code, Codex, Cursor and other agents, and runs on Node.js or Bun. Documentation is at skills.sh.

npx skills add nickture/skills    # Node.js
bunx skills add nickture/skills   # Bun

The command asks which skills to install and into which agents. The --skill flag installs one skill, and -g installs it for all projects.

npx skills add nickture/skills --skill nickture-interface -g
bunx skills add nickture/skills --skill nickture-interface -g

As a Claude Code plugin

/plugin marketplace add nickture/skills
/plugin install nickture-skills@nickture

The first command adds the plugin marketplace, the second installs the plugin with both skills. To install just one skill, use the skills CLI with --skill.

Then turn on auto-update. In /plugin, open the Marketplaces tab, select nickture and select Enable auto-update. Claude Code keeps it off for marketplaces outside Anthropic’s, and without it the plugin stays on the version you installed. With it on, each version comes from GitHub as soon as it is released, without a review.

Manually

Copy the skill folder into the agent’s skills directory.

If you symlink a clone of the repository instead of copying, git pull updates the skill.

git clone https://github.com/nickture/skills.git ~/nickture-skills
mkdir -p ~/.claude/skills
ln -s ~/nickture-skills/skills/nickture-interface ~/.claude/skills/nickture-interface

Usage

A skill loads on its own when the task matches its description. To call it explicitly, type /nickture-interface in Claude Code and Cursor or $nickture-interface in Codex.

Example requests:

• “Check the checkout screen against nickture-interface”
• “Rewrite the error messages in the sign-up form with nickture-text-ru”
• “Run the nickture-interface quick pass on the pricing page”

A review of a finished project reads every rule. An edit, such as a new component or a changed screen, reads the slop sections and only the sections it touches. The agent picks them from the section table in SKILL.md.

Run a review in plan mode. In it the agent reads the rules, lists the ones that are broken and proposes fixes. It changes files only after you approve the plan, so you can drop a fix you disagree with before it reaches the code.

A weaker model or a low reasoning effort reads selectively and skips rules, so use the strongest model you have. In Claude Code that is Opus or Sonnet with effort set to high or max, and some of them default to medium.

/effort high

max lasts until the end of the session, high carries over to the next ones. In Codex, choose the model and