Agent Skills (MengTo/Skills)
github.com/mengto/skills- Category
- AI Agents
- Rank
- No. 2055Tools index
Previous survey · No. 2065 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Design & Media
- Interfaces
- Agent Skill / Plugin
- Builder
- MengTo
- GitHub
- 6.7k stars
- Date
About
A curated collection of agent skills aimed at designers and builders working with Codex, Claude, Cursor, and other AI coding agents. The skills cover rich user interfaces, playable games, frontend systems, agent loops, automations, and reusable workflows; the flagship web-design workflow turns a screen recording into a detailed prompt an LLM can use to one-shot HTML, and a companion skill converts an existing HTML page into reusable interaction prompts.
What it does
A library of plain Markdown instruction packs that a coding agent reads before it starts a task. Each pack is a folder holding one required instructions file with steps, defaults and pitfalls, plus optional reference links, long-form articles, helper scripts and a standalone demo page. You copy a folder into your agent's skills directory or open the file as context. The README counts 157 skills in eight groups, most of them about web visuals: scroll motion, shaders, 3D scenes, browser game systems, page layouts, plus a handful of capture and shipping workflows.
Why it's ranked here
The format is the selling point: no runtime, no package, no lock-in. Any agent that can read Markdown can use it, and the README gives loading instructions for Codex, Claude Code and Cursor. It is MIT licensed. The repository also polices its own demos with a validation script that rejects thin prompt recipes, oversized previews and pages that call remote URLs. That is more quality control than most prompt collections carry, and it is why this ranks above a typical folder of saved prompts.
What's good
The demo checks have real numbers behind them. Every skill demo needs a standalone page with a doctype, viewport tag and title, a recreation prompt of at least 120 words with three fixed headings, and a preview screenshot at least 1200 by 700 pixels and under 2 MB. Non-workflow demos must include a reduced-motion fallback. Demos imported from the Neuform design source run inside a sandboxed frame that allows scripts but not same-origin access, behind a content security policy, with checksums that a test script verifies. The 3D and workflow guides each include a table telling you which skill to reach for.
Tradeoffs
Parts of the repository still read as one person's working folder. The web-design notes call themselves an internal draft pending finalization, and the agent guide points at the author's own machine path. Some skills are personal, such as one that drafts X posts in the author's voice. The counts drift: the Codex section is headed nineteen but lists sixteen, and the 3D guide describes fifteen skills while the README tree shows nine. Codex is the first-class target; the README tells agents to use the Codex browser for browser work in the repository. There is no installer, so updates mean recopying folders by hand.
How to use it well
Pick the narrowest skill for the job rather than loading the whole library, which is what the 3D guide itself advises. It suits a designer or front-end developer who wants an agent to reproduce a specific effect: a gradient border, progressive blur, a scroll reveal, a globe, a water surface. Open the skill's demo page and recreation prompt first to judge whether the output matches your taste. It does not cover backend, data or testing work beyond its own browser-game QA and screenshot workflows.
Technical notes+
README.md defines the folder contract: one required skill file with frontmatter and workflow, an optional links-only references file, optional article, assets, scripts and a demo folder holding a standalone page and recreation prompt. scripts/validate-skill-demos.mjs enumerates tracked skill files via git ls-files and enforces per-demo rules: doctype, viewport, title, one h1 and a main landmark for non-imported demos, a prefers-reduced-motion check outside the codex category, no remote URLs, no absolute personal paths or email addresses, inline scripts parsed with node:vm, JPEG preview dimensions read from the file header with 2 percent drift tolerance, assets capped at 5 MB, and demo index row counts matching the skill count. scripts/test-sync-neuform-security.mjs asserts private-IP blocking, HTTPS-only allowlisted hosts, query-string stripping, an allow-scripts sandbox without allow-same-origin, and sha256 checksums across 38 source manifests. Both scripts use only Node built-ins. CLAUDE.md hardcodes the author's local machine path.
Observed
- License
- MIT
- Format
- Plain Markdown skill folders, one required instructions file per skill
- Install surface
- No package; copy a skill folder into the agent's skills directory or load it as context
- Agents documented
- Codex, Claude Code, Cursor
- Demo validation
- Node script with built-in modules only, enforcing demo page, prompt and preview rules
- Imported demo isolation
- Sandboxed iframe with scripts allowed and same-origin denied, plus content security policy
- Draft status
- Web-design folder notes describe it as an internal draft pending finalization
Read from README.md, LICENSE, CLAUDE.md, agent-skills/web-design/README.md, agent-skills/web-design/WEB-DESIGN-SKILLS.md, agent-skills/workflow/README.md, agent-skills/3d/README.md, scripts/validate-skill-demos.mjs, scripts/test-sync-neuform-security.mjs.
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.