How to Install Agent Skills in Claude, ChatGPT, Cursor, Kimi & GLM

Installing Agent Skills means placing a folder with SKILL.md where your agent discovers skills — then verifying Claude, ChatGPT, Cursor, Codex, Gemini, Kimi, or GLM can trigger it from the description. For design work, install one taste skill, add a DESIGN.md from the gallery, and run the same UI prompt with the skill on. Paths differ by tool; the file format does not.

Before you install

You need:

A skill folder containing SKILL.md (YAML frontmatter + Markdown body)
An agent that supports the Agent Skills standard (Claude Code, Cursor, Codex / ChatGPT coding, Gemini CLI, and many others — including stacks that run Kimi or GLM)
Optional: a design template so the skill has tokens to obey

If you are new to the format, read What are Agent Skills? first.

Install paths by agent

Agent	Project (repo) skills	User (global) skills
Claude Code	`.claude/skills/<name>/SKILL.md`	`~/.claude/skills/<name>/SKILL.md`
Cursor	`.cursor/skills/<name>/SKILL.md`	Also discovers `.claude/skills/` and Codex-compatible dirs
ChatGPT / Codex	`.agents/skills/` or `.codex/skills/`	`~/.agents/skills/` (and related Codex user paths)
Gemini CLI	`.gemini/skills/`	`~/.gemini/skills/`
Kimi / GLM CLIs	Vendor skills directory or shared Agent Skills path	Check the product docs for the exact folder

Team tip: Commit project skills to git so Claude, Cursor, and ChatGPT sessions on the same repo share craft rules. Keep personal experiments in the user-level folder.

Method A — Copy a skill into the repo (most reliable)

Create the folder: .cursor/skills/tasteful-ui/ (or .claude/skills/…)
Add SKILL.md with name + description + steps
Restart the agent session or reload the window so discovery refreshes
Ask: "What skills do you have?" or invoke /tasteful-ui if your client supports slash skills
Run a design task and confirm the skill body influenced the plan

Example frontmatter:

---
name: tasteful-ui
description: Use when building or redesigning UI, landing pages, or marketing sections. Applies anti-slop layout and type rules.
---

Method B — Install from a registry or CLI

Many ecosystems expose npx skills add … or vendor-specific installers. Flow:

Find the skill on a registry or on /skills
Run the documented install command for your agent
Confirm files landed under a discovered skills path
Open Claude Code, Cursor, or ChatGPT/Codex and trigger a matching task

Prefer design skills with demos. A popular download with no craft rules still ships slop on Claude and on Kimi.

Method C — One skill, multiple agents (monorepo pattern)

If your team mixes tools:

repo/
  .agents/skills/tasteful-ui/SKILL.md   # Codex / ChatGPT-friendly
  .claude/skills/tasteful-ui -> ../../.agents/skills/tasteful-ui
  .cursor/skills/tasteful-ui -> ../../.agents/skills/tasteful-ui
  DESIGN.md

Symlinks keep a single source of truth. Cursor often already reads Claude/Codex skill directories — check your Cursor version docs before duplicating.

Verify the skill actually loads

Claude

Start Claude Code in the repo
Ask it to list available skills or to use /skill-name
Give a UI task that matches the description
Confirm the response references skill steps (anti-patterns, DESIGN.md, section scope)

ChatGPT / Codex

Open a coding session with repo access
Mention the skill or rely on auto-trigger from the description
Watch for plan steps that mirror the skill body
If nothing fires, invoke explicitly with the client's skill mention syntax

Cursor

Agent mode / Composer with the project open
Use /skill-name when auto-invocation is off
@-reference DESIGN.md alongside the skill for UI tasks
Keep diffs scoped — skills help most when you do not ask for the whole app at once

Kimi, GLM, Gemini

Confirm the CLI or IDE plugin supports Agent Skills paths
Install to the documented folder
Run the same landing-page prompt you use with Claude or ChatGPT
Compare: skill off vs skill on — the delta should be obvious

Model quality still matters. A strong Claude or ChatGPT model follows nuanced craft better than a weak local run — but without a skill, every model regresses to generic UI.

Design-skill install checklist

Use this when the goal is tasteful UI, not generic coding helpers:

Skill description mentions UI / landing / redesign triggers
Body bans specific slop patterns
Body tells the agent to read DESIGN.md when present
Repo has a DESIGN.md from Agent's Design or your brand
Only one primary taste skill enabled
Tested on your daily driver (Claude or Cursor or ChatGPT) and spot-checked on a second model (Kimi, GLM, or Gemini)

Troubleshooting

Problem	Fix
Skill never triggers	Sharpen `description`; invoke manually; restart session
Skill triggers on everything	Narrow description; set manual-only if supported
Output still looks templated	Add DESIGN.md + gallery template; see stop AI UI looking templated
Works in Claude, missing in Cursor	Copy or symlink into a path Cursor discovers
Works in ChatGPT, fails in Kimi/GLM	Confirm that client’s skills directory and reload
Overlapping skills fight	Disable extras; keep one craft skill

Minimal DESIGN.md hook (paste into skill body)

## Design constraints

1. If DESIGN.md exists, read it before writing UI. Do not invent tokens.
2. Build one section per request.
3. No purple gradients, no Inter unless specified, no three identical feature cards.
4. Prefer existing components in components/ui when present.

That block alone improves Claude, ChatGPT, Cursor, Kimi, and GLM sessions.

After install — first session script

Browse a skill demo on /skills
Install the skill via Method A or B
Copy a template DESIGN.md from the gallery
Prompt: "Using the tasteful-ui skill and DESIGN.md, implement the hero only"
Review in the browser; then next section
Repeat in a second agent (e.g. Claude plan → Cursor apply, or ChatGPT → local GLM check)

Related guides

Skills are files. Install them where your agent looks, keep DESIGN.md beside them, and your Claude, ChatGPT, Kimi, and GLM sessions stop negotiating taste from scratch every time.