MengTo/Skills: Portable Agent Skills for Codex, Claude and Cursor

The problem MengTo/Skills solves for agent users

Most people drive an AI coding agent with chat history. The good instruction you worked out last Tuesday lives in a scrollback buffer, and the next session starts from nothing. MengTo/Skills takes the position that a prompt is an asset: if it worked once, it should be a file. The README states this directly under a section called Prompts are assets, with the sub-points store prompts as files, version them, and build libraries plus stylecards.

The audience is narrow and specific. The description names designers and builders using Codex, Claude, Cursor and other AI coding agents. The skills themselves lean toward visual output: landing pages, prompts from screen recordings, 14 SVG illustration styles, Three.js, and a 3D group covering virtual tours, sky rays, falling leaves, four seasons and Retina rendering at 200%. A second cluster covers game development, with a game-combat category and reusable game architecture and gameplay QA.

This is not a runtime library. Nothing here executes on its own. The unit of delivery is a Markdown file an agent reads before it acts, which means the value shows up only when your agent is configured to look at it.

How a SKILL.md folder is structured and loaded

The mechanism is a folder contract, not an API. The README spells it out as agent-skills/<category>/<skill-name>/ with SKILL.md required and REFERENCES.md, ARTICLE.md, assets/, scripts/ and demo/ optional. A visual or interaction skill may add demo/index.html and demo/PROMPT.md; a workflow skill may add fictional demo/input.md and demo/expected-output.md handoffs.

The split between SKILL.md and REFERENCES.md is the part worth copying. The README says REFERENCES.md is links only, with the stated reason that this keeps SKILL.md lean. That is a deliberate constraint on context budget: the procedural file stays short, and the agent pulls external material only when it needs it. The README also asks that skills stay procedural, with steps, patterns and guardrails rather than encyclopedic coverage, and that they prefer explicit triggers, quoting Use when... as better than vague descriptions.

Agent support is described per tool. Codex should load the relevant SKILL.md before acting, with repo behavior in AGENTS.md. Claude Code can reference a skill from CLAUDE.md, copy it into a Claude skills setup, or open the SKILL.md directly as working context. Cursor should point rules or chat context at a specific skill folder. Other agents follow the same contract: read the narrowest matching skill first, then its steps and linked references.

The README gives a command for finding what exists rather than trusting a list:

find agent-skills -name SKILL.md | sort

That command is the source of truth the README points at, which matters because the count in prose and the tree in the repository can drift apart.

Installing MengTo/Skills and running a first workflow

There is no package to install. The repository is the distribution. Clone it, then point your agent at one folder.

git clone https://github.com/MengTo/Skills.git
cd Skills
find agent-skills -name SKILL.md | sort

The third command prints every skill path. Pick one narrow skill rather than a category. For a first run, the flagship sequence in the README starts with Video to Super Prompt at agent-skills/codex/video-to-superprompt/SKILL.md, which the README describes as turning a screen recording of a design, landing page or animation into a detailed prompt that Fable 5 can use to one-shot HTML.

For Claude Code, the README gives three options: reference the skill from CLAUDE.md, copy it into a Claude skills setup, or open the SKILL.md directly as working context. The lightest of those is the last one, and it is also the easiest to verify.

cat agent-skills/codex/video-to-superprompt/SKILL.md

Read it before you feed it to an agent. You should see frontmatter followed by a workflow, and the workflow should state when to use the skill and what to do first. If it reads like an article instead of a procedure, that folder is not the one to start with. For Cursor, the README's instruction is to point rules or chat context at the specific skill folder rather than the whole repository, which keeps unrelated skills out of the context window.

Skills that ship a demo folder are the ones you can check without trusting prose. The README describes demo/index.html as standalone HTML, CSS and JavaScript, and demo/PROMPT.md as exact recreation and remix prompts. Open the HTML file in a browser, then compare what it renders against what PROMPT.md asks for. That is the closest thing this project has to a test.

Where this library breaks down

The repository is plain Markdown and folders, which the README calls intentional. That choice has costs. A skill can be well written, thin, or stale, so read it before relying on it. The README's own guidance, keep skills procedural and prefer explicit triggers, is advice rather than enforcement.

Portability is the second soft edge. The README says each skill should work for any user, repo or workspace unless the user supplies project-specific context through local agent instructions. That is a promise about design intent, not a guarantee about output. A skill that names Tailwind, GSAP or Three.js assumes that stack. Drop a Three.js skill into a project that renders with something else and the agent will follow the file anyway.

The third limitation is agent coupling. Codex, Claude Code and Cursor are each described differently because each loads context differently. A skill that works when opened directly as context in Claude Code may behave differently when referenced from a Cursor rules file.

Finally, there is no release history in the repository. The repository has no published GitHub releases, so there are no version tags to pin against. If you want a fixed, auditable set of instructions, you would pin a commit hash yourself.

MengTo/Skills compared with a prompt library or a framework

The nearest alternative for most teams is a shared prompt document: a Notion page, a wiki, or a prompts/ directory of loose text files. The difference is the folder contract. A prompt document gives you text; this gives you a directory with a required procedural file, an optional links file, and optional demo and script folders. That structure is what lets an agent read the narrowest matching skill first, which the README names as the contract for agents it does not explicitly support. A flat prompt file has no equivalent notion of scope.

The other alternative is a framework with a runtime, where instructions ship as code with a version number and a test suite. That trade runs the other way. You get pinning and verification, and you give up the ability to edit the instruction in a text editor and have the agent pick it up on the next read. MengTo/Skills sits firmly on the text side. The README's stated goal is to turn prompts, workflows, style systems, capture recipes and debugging habits into versioned files an agent can load and follow, and versioned here means tracked in git, not published as a numbered release.

Licence, maintenance and upgrade cost

The repository is MIT licensed, with a LICENSE file at the top level. For a collection of Markdown instructions, that is permissive: you can copy a skill folder into your own repository, adapt it, and keep your changes private. The practical implication is that skills you adapt stop receiving upstream edits unless you track them yourself. This is not legal advice; read the LICENSE file and your own obligations.

On maintenance, the last push was on 2026-11-02, which is recent. The README describes the repository as a snapshot containing 176 skills across nine categories, and that word matters. A snapshot number in prose goes stale as folders are added, which is why the README points at find agent-skills -name SKILL.md | sort instead. Treat the count as approximate.

Upgrading is a git pull, and the cost is review rather than migration. There is no version number to diff against, so a pull can change a skill you depend on without any signal beyond the commit. If a skill folder has become part of your workflow, copying it into your own repository and diverging is a defensible move under MIT, and it removes the review burden at the price of upstream fixes.

Editorial conclusion

Adopt this if you already drive Codex, Claude Code or Cursor and you keep rewriting the same design and Three.js instructions in chat. It is the wrong choice if you need a package manager install, a versioned release, or skills that run without an agent reading them. Before you commit, open one SKILL.md from agent-skills/codex/ and one demo/PROMPT.md, and check that the trigger language in the frontmatter matches how you actually phrase requests. Also read the folder contract in the README, because the repository distinguishes SKILL.md, REFERENCES.md, ARTICLE.md, scripts/ and demo/ and the value of a skill depends on which of those it ships.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. MengTo/Skills on GitHub
4. README