MengTo/Skills has no tagged releases — pin the commit

MengTo/Skills is a public, MIT-licensed folder of file-based "Agent Skills" — but the star metric everyone quotes, the skill count, keeps moving. Before you clone it, it helps to know exactly what is in the repo today and why the tally floats.

What MengTo/Skills actually contains: 78 MIT modules and a floating tally

MengTo/Skills is a public GitHub repository of reusable, file-based prompt and workflow packages built by Meng To, the designer behind Design+Code. As of July 2026 the live README reports 78 skill folders — even though circulated figures of 95 (marketing snapshots) and 75 (launch-era counts) still float around. The count is unstable by design, so the README names its own source of truth: run find agent-skills -name SKILL.md | sort against a cloned copy to get an authoritative number for the commit you actually have.

Quick Answer: MengTo/Skills is an MIT-licensed GitHub repo of file-based Agent Skills. The live README shows 78 skill folders as of July 2026; ignore the circulated 95 and 75 figures and count them yourself with find agent-skills -name SKILL.md | sort.

Those 78 modules split across four categories, heavily weighted toward visual web work:

Category	Skills	Representative content
Web design	63	GSAP, ScrollTrigger, Three.js/WebGL, Tailwind, Lenis, Cobe, Vanta, Matter.js, glass UI, skeuomorphic UI
Codex workflows	12	video-to-superprompt, html-to-interaction-prompts, stitched-full-page-capture
Media sourcing	2	Image/asset acquisition helpers
UI prompting	1	design-first-ui-prompting system

The repo state at research time: roughly 2,500 stars, 271 forks, 85 commits, 2 open issues, and 0 tagged releases under an MIT license (MengTo/Skills). That zero-releases detail matters later: with no tags, reproducible installs depend on pinning a commit.

Each module follows the emerging Agent Skills convention: a required SKILL.md holding YAML frontmatter (name, description) plus a step-by-step body, with optional REFERENCES.md, ARTICLE.md, and a demo/ subdirectory of standalone HTML/CSS/JS (agent-skills tree). Because the folders already match this layout, they drop into Claude Code, Cursor, and Codex without conversion.

Getting a reproducible copy: npx vs. manual directory placement

There are two ways to pull the modules into your agent: an automated CLI install, or manual directory placement. The fastest path is the community skills npm CLI, which deposits each folder into the agent's expected skills directory for you. At crawl time the package sat at version 1.5.19, with 86 published versions and 58 dependents, and supports Claude Code, Codex, Cursor, OpenCode, and Kiro among others . A single command targets multiple agents at once:

npx skills add MengTo/Skills --skill '*' --agent claude-code --agent cursor

The manual route gives you tighter control over which of the 78 folders land where. Clone the repo, then copy or symlink the folders you want. For Claude Code, project-scoped skills live at .claude/skills/<name>/SKILL.md and personal skills at ~/.claude/skills/<name>/SKILL.md . Cursor adopted the same SKILL.md layout in version 2.4, shipped January 22, 2026, so the folders drop into either tool without conversion .

The reproducibility catch is that MengTo/Skills has no tagged releases . Whichever install path you pick, capture the commit SHA immediately after cloning so teammates can rebuild the exact folder set later:

git -C MengTo-Skills rev-parse HEAD

Store that hash in your dotfiles or the project README. The illustrative Python below (not executed here — it needs network access to api.github.com) shows the same logic programmatically: check for tags, and when there are none, print a pinnable commit reference.

import json
import urllib.request

repo = "MengTo/Skills"
api = f"https://api.github.com/repos/{repo}"
headers = {"Accept": "application/vnd.github+json", "User-Agent": "pin-commit-demo"}


def get(path):
    req = urllib.request.Request(api + path, headers=headers)
    with urllib.request.urlopen(req, timeout=20) as response:
        return json.load(response)


info = get("")
tags = get("/tags")
branch = get(f"/branches/{info['default_branch']}")
sha = branch["commit"]["sha"]

print(f"{repo} tags: {len(tags)}")
if tags:
    print(f"tag available: {tags[0]['name']}")
else:
    print("no tags; pin the commit")
    print(f"git+https://github.com/{repo}.git@{sha}")

Decide placement by intent. Personal design taste — animation style, motion preferences, visual defaults — belongs under the user-level ~/.claude/skills folder, where it follows you across projects. Shared team conventions belong in .claude/skills committed to the project repo, so everyone resolves the same skills from the same pinned commit .

Reproducibility concerns: YAML divergence and no lockfile

The pin matters because MengTo/Skills ships no lockfile and no tagged releases, so two teammates cloning main on different days can resolve different folder sets. The catalog tally moved at least three times during research — a circulated 95, a launch-time 75, and the current README's 78 — which is exactly the drift a commit SHA neutralizes. Treat the SHA as the functional equivalent of a lockfile: record it, commit it, and re-audit the delta with find agent-skills -name SKILL.md | sort before you bump it.

The second gap is YAML portability. The SKILL.md frontmatter is baseline-compatible everywhere, but platform-specific keys are not universally honored. Claude Code reads fields like context: fork, allowed-tools, and disable-model-invocation; Cursor — which added Agent Skills in version 2.4 on January 22, 2026 — and Codex silently ignore them. No error is thrown. A skill that relies on context: fork for isolation simply runs inline in the wrong host, so a workflow that behaved on one machine can quietly misbehave on another with no diagnostic.

The repo's own README frames dedicated per-host install notes as a maintenance idea, not a stabilized official setup, so the burden of validation sits with you [repo]. Before committing anything to shared dotfiles, load each folder in the actual target agent and confirm invocation, tool access, and any fork/isolation behavior resolve as expected. Validate first, pin the commit, then propagate — reversing that order is how silent no-ops reach production.

Noteworthy MIT modules: motion, interaction, and landing-builder packs

Once your copy is validated and pinned, a handful of modules justify the effort on their own. The most technically demanding is video-to-superprompt, the flagship Codex workflow: it tells the agent to locate a video, run ffprobe, extract frames with ffmpeg, then analyze story, layout, motion, visual, and technical layers before naming concrete motion mechanisms — pinned sections, masks, shader fields, scrubbed timelines, and reduced-motion fallbacks — and verifying paths and screenshots before it finalizes a reusable prompt pack . It is less a prompt than an operational pipeline.

Three more packs are worth pinning:

• landing-page — collects one primary CTA, offer, ICP, objections, proof, traffic source, brand voice, and mobile constraints, then outputs a page outline, hero copy, benefits, how-it-works, FAQ, and an SEO/AEO layout recommendation . It is the clearest expression of the repo's "prompts as assets" stance: store prompts as versioned files and build them into libraries.
• design-first-ui-prompting — formalizes the "prompt like a design system" approach: specify goal, format, margins, grid, hierarchy, type, colors, materials, imagery, exact copy, and negative prompts, then iterate by changing one variable at a time rather than rerolling blindly .
• html-to-interaction-prompts — decomposes existing HTML into reusable prompts for sections, hover states, animations, buttons, and WebGL effects, which is useful for reverse-engineering a live UI into agent-ready instructions .

Meng To frames the underlying method plainly: "screenshots convey fonts, spacing, colors, layout rhythm, and icon style better than paragraphs" — Meng To, Design+Code (source: MengTo/Skills). The takeaway: pin a commit, start with these four modules, and adapt the frontmatter to your target agent — the packs are concrete enough to use as-is, but the repo carries no releases, so reproducibility is still your responsibility.

Frequently asked questions

How many agent skills does MengTo/Skills actually contain?

As of July 2026, the live README reports 78 skills across four categories . Circulating figures of 95 and 75 are outdated directory snapshots. The count moves through direct commits, so don't trust any quoted number — clone the repo and run find agent-skills -name SKILL.md | sort, which the README itself names as the source of truth.

Why does MengTo/Skills have no tagged releases?

The repo has grown entirely through direct commits with no formal versioning — roughly 85 commits and 0 release tags at research time . The SKILL.md file format is stable, but the folder inventory is not versioned. That gap is why reproducible installs require pinning a specific commit SHA rather than relying on a tag.

How do I install MengTo/Skills into Claude Code?

Two paths work. The fastest is the skills npm CLI: npx skills add MengTo/Skills --skill '*' --agent claude-code . Alternatively, copy selected folders manually to .claude/skills/<name>/SKILL.md (project scope) or ~/.claude/skills/<name>/SKILL.md (personal scope). Because there are no releases, capture the git SHA immediately so your install is reproducible.

Will these skill folders work in Cursor and Codex?

Yes, at the baseline SKILL.md level. Cursor added Agent Skills in version 2.4, released January 22, 2026 , and Codex supports the convention, so folders drop in without conversion. But both silently ignore Claude-specific YAML fields like context: fork and allowed-tools. Test your target agent before depending on those fields.

What is the Agent Skills format and who defined it?

Agent Skills is an open convention formalized by Anthropic: a directory containing a required SKILL.md file with YAML frontmatter (name, description) plus a step-by-step instruction body . It relies on progressive disclosure — agents load names and descriptions at startup, full instructions only on activation, and supporting resources only as needed. Optional companions include REFERENCES.md, ARTICLE.md, and a demo/ folder.