Skills that make agents do the work.
Drop-in skills for AI agents, in one place: build slide decks, ship code, design, run research — sorted by the job they do.
Small, Hackable Skills for Real Engineering
The skills Matt Pocock uses daily: get grilled until nothing is ambiguous, turn talk into a spec and linked tickets, build red-green-refactor, run hard bugs through a fixed loop. Small and hackable, not process-owning.
205.2K stars
Four Karpathy Rules for Coding Agents
One 65-line CLAUDE.md that turns four coding-agent habits Karpathy called out into checkable rules: no guessing on your behalf, no over-building, no drive-by edits, and success criteria defined before any code.
200K stars
shadcn/ui: Components You Copy, Then Own
Not an npm dependency — the component source lands in your repo and becomes yours to edit. The project's own framing: start here, then make it your own component library.
120.6K stars
UI UX Pro Max: A Design System Generated From Your Brief
Give it the project context and its reasoning rules pick the page structure, UI style, palette, font pairing, and key effects — handing back a formed design system instead of a blank canvas.
113.9K stars
Output Compression: 65% Is Prose, Not the Whole Story
Six intensity levels, three of them classical-Chinese compression. A dedicated document says it plainly: 65% is the average over ten chat-style prompts, while a third-party run of 86 coding tasks measured 8.5%.
96.2K stars
Open Design: A Design Engine That Runs on Your Machine
It turns the Claude Design loop — pin the brief, lock the direction, stream the artifact, critique, deliver — into files on disk: skills, templates, design systems and plugins your local coding agents can read and rewrite.
84K stars
24 Engineering Skills, /spec Through /ship
Eight slash commands along the development lifecycle — spec, plan, build a slice at a time, test, review, webperf, simplify, ship — plus skills that self-activate: API work pulls in interface design, UI work in frontend.
82K stars
Understand Anything: Read a Codebase as a Graph
A multi-agent pipeline sweeps the project, builds a knowledge graph of every file, function, class and dependency, then hands you a dashboard you can click, search and question. Knowledge bases and docs work too.
77.7K stars
Taste Skill: The Anti-Slop Layer for AI-Built UI
Portable agent skills aimed at the sameness of AI-generated interfaces — stronger layout, typography, motion and spacing. A separate set generates reference imagery only, no code.
72.6K stars
career-ops: Filter the Listings Before You Apply
Paste a job link and it scores the role one to five across five weighted dimensions, plus a separate read on whether the posting is real at all. It is a filter, not a blaster: below four, it tells you not to apply.
63K stars
last30days: Search People, Not Editors
Give it a topic, company or person and it searches twenty-odd platforms in parallel for the last thirty days, ranks by upvotes, likes and real-money odds, then has an agent judge synthesise one brief.
57.4K stars
Impeccable: 59 Design Checks That Need No LLM
One skill plus a CLI and a browser extension behind a single entry point, with 23 commands. The 59 detection rules are deterministic — running them needs no LLM and no API key.
55.9K stars
Remotion: Write Video in React
The framework describes frames in React; twelve companion skills teach an agent how to drive it. Those skills install separately — they do not ship inside the framework.
55.6K stars
Writing Obsidian's Own File Formats
Five skills that teach an agent Obsidian's native formats directly: wikilinks and callouts in Markdown, views and formulas in .base files, nodes and edges in JSON Canvas, plus the official CLI and a web-extraction tool.
44.2K stars
49 Marketing Skills on One Shared Foundation
Conversion, copy, SEO, paid distribution, retention, growth engineering and RevOps are the groups. What matters is the foundation: every other skill reads product-marketing first for the product, audience and positioning.
43.2K stars
An Academic Pipeline Gated by Two Integrity Checks
Four skills chained into ten stages: research, write, review, revise, finalise. An integrity gate sits after writing and again before finalising, each running a seven-mode blocking checklist that stops the draft cold.
41.1K stars
agent-browser: A Browser CLI Built for Agents
A native Rust CLI with a resident daemon driving Chrome over CDP. The body of it is deterministic commands, not natural-language control — that lives in one subcommand and needs a gateway key configured first.
40K stars
HyperFrames: Write HTML, Render MP4
It renders HTML, CSS, media and seekable animations into deterministic MP4. Deterministic is the point: the same input always yields the same frames, so you can seek, preview and re-render safely.
39.6K stars
158 Research Skills, and Advice Not to Install All
Covering bioinformatics, cheminformatics, mass-spec proteomics, clinical pharmacology, imaging, geospatial, materials and astronomy; one database skill gives deterministic, sourced access to 78 public repositories.
32.8K stars
A Workspace CLI That Reads the API at Runtime
A single Rust binary with no built-in command table: at runtime it reads Google's Discovery service and builds the command tree, cached 24 hours. The README states twice that it is not an official Google product.
30.2K stars
Nine Frontend Skills From the Vercel Team
The React performance skill carries 40-plus rules in eight impact-ranked groups, the design skill reviews UI against 100-plus, the writing skill reviews docs against 80-plus, and one audits your Vercel bill.
29.8K stars
Emil's Nine Design-Engineering Skills
Design-engineering experience from Vercel and Linear, split into nine skills: build, review and improve motion, then sharpen vocabulary, Apple-inspired principles, library choice and multi-option prototyping.
27.4K stars
Frontend Slides: Web Decks for People Who Don’t Design
It shows rather than asks — instead of making you describe your aesthetic in words, it generates visual previews and lets you pick one, then builds from there. It also converts existing PowerPoint files to web.
27K stars
A PM Marketplace: 9 Plugins, 68 Skills
Not a plain skill collection but a marketplace. Skills load by context, 42 commands chain several skills into end-to-end flows, and the plugins group by product-management domain.
24.9K stars
Baoyu’s Skills for Visual Content
Twenty-odd skills across content generation, AI generation, and everyday utilities. The centrepiece turns an article into a set of one to ten social cards, with visual style and information layout as two independent dials.
24.6K stars
SuperClaude: 30 Commands Behind an sc: Prefix
It injects behavioural instructions and slash commands into Claude Code, and after install every command carries the sc: prefix — typing /implement from the old docs finds nothing. Seven modes, twenty agents.
23.7K stars
Guizang Web Decks: Two Locked Visual Systems
Turn an article or Markdown into a single-file HTML deck that pages sideways, with illustrations and covers. Two visual systems picked by content: magazine editorial for narrative, Swiss International for fact and method.
23.3K stars
Six Skills Khazix Uses Himself
Six skills, one job each: break down a goal, clean up disk space, write research reports, track what is trending, tidy files, and write. From Khazix, founder of Xushi Media.
19.3K stars
Lark's Official CLI, in Three Command Tiers
Plus-prefixed shortcuts serve humans and agents alike, with smart defaults and dry-run; 100-plus API commands map one-to-one onto platform endpoints; below that, lark-cli api reaches 2,500-plus endpoints raw.
16.2K stars
Stop Slop: Strip the AI Tells Out of Prose
AI writing has fixed phrases, structures and rhythms. This turns them into a checklist you can work through line by line, then scores the result across five dimensions.
15.2K stars
Claude SEO: 25 Sub-Skills Auditing in Parallel
Technical SEO, content quality, Schema markup, AI-search optimisation, local and international SEO each go to a specialist agent; they run in parallel and return one prioritised action plan.
13.4K stars
GSAP's Own Eight Animation Skills
Split by topic: core tweens, timelines, ScrollTrigger, plugins, utilities, React, performance and framework integration, with an llms.txt index written for agents.
13.1K stars
Kami: Documents That Read Like Printed Pages
The author's diagnosis: AI already writes good documents; what's missing isn't capability but constraint. Without a design system, every session drifts into generic grey. Kami supplies the constraint.
10.4K stars
29 Skills and a Router That Picks One Per Turn
The routing is the point: it reads context and selects exactly one skill per turn, then offers three ways forward — back to the router, into a named skill, or stay where you are.
9.3K stars
Xiaohei: A Hand-Drawn Illustration Set for Articles
Note this is a Codex skill, not a Claude one. The style is fixed: pure white ground, thin black hand-drawn line, Chinese annotations handwritten in red, orange and blue, subject at 40 to 60 percent of the frame.
9.1K stars
Giving an Agent Real Web Access
Tool selection splits four ways — search, fetch, curl, and a CDP-driven browser — with Jina as optional pre-processing. It works with Cursor, Gemini CLI and Codex CLI, not only Claude Code.
8.6K stars
HTML Anything: An Agent-Era HTML Editor, Local
Its argument: Markdown is the draft, HTML is what people actually read. No API key — it reuses the coding-CLI session you are already logged into and turns the draft into something publishable, locally.
8.1K stars
Stitch Design Skills, Behind an MCP Gate
Three plugins bundle fifteen skills covering code-to-design, static HTML extraction, a five-page site from one sentence, and walkthrough video. The hard prerequisite is a registered Stitch MCP server.
7.9K stars
Waza: Eight Engineering Habits, Split by Stage
Eight compact skills cover planning, UI, debugging, review, reading, writing, research and agent health. Each has one trigger and one stopping point, ready for Codex, Claude Code and other hosts.
6.8K stars
Summarize: A Sidebar and a CLI
Not a skill but a complete product — an npm package plus a browser extension, with sidebars in both Chrome and Firefox. Web pages, PDFs, images and media each take their own extraction path.
6.5K stars
Trail of Bits: 39 Security Plugins
Organised into nine categories — smart-contract security, code audit, malware analysis, verification, reverse engineering, mobile — the 39 plugins hold 77 SKILL.md files, code audit alone accounting for 19.
6.5K stars
70 Product-Management Skills in Three Layers
The workflow layer runs day-to-week processes, the interactive layer asks three to five context questions before offering numbered options, and the component layer holds 30-to-90-minute templates.
6.3K stars
A Daily Brief on What AI Builders Are Doing
The sources are a fixed, named list: six podcasts, twenty-six builders on X, two official blogs. Fetching happens centrally and refreshes daily, so your side makes a single request.
6.2K stars
14 n8n Skills Plus a Layer of Hooks
The hard prerequisite is the n8n-mcp server: without it the skills are largely inert. Alongside the 14 capability skills sit a resident router skill and three hooks that pull them into context at the right moment.
6K stars
Social Cards: 28 Layouts, No Custom Colours
Three canvases are fixed: 1080×1440 for Xiaohongshu, 21:9 for WeChat, 1080 square. Output is a PNG captured through Playwright; the Live Photo branch produces a JPG, MOV and .pvt set instead.
5.9K stars
Anthony Fu's Nineteen Frontend Skills
Three tiers: two hand-written, nine generated from official documentation, eight vendored in from upstream repos. The mechanism is git submodules pointing at upstream doc sources, so they track upstream.
5.7K stars
Turning a Codebase Into a Clickable Course
The output is a directory, not one file: prebuilt CSS and JS, one HTML per module, assembled into an index by a build script. The only external dependency is a font CDN.
5.3K stars
BrowserAct: Three Browser Modes and a Human Handover
A Python CLI with three modes: reuse your local Chrome session, a privacy mode with a fresh fingerprint and rotating proxy each time, or a fixed identity for running several accounts in parallel.
5.2K stars
SwiftUI Code Review, Not Code Generation
It positions itself as a reviewer: nine steps, each loading its own reference file, from deprecated APIs through view structure, data flow and navigation to accessibility, performance and hygiene.
4.4K stars
Remotion's Own Twelve Video Skills
One each for creating a project, writing markup, driving the Studio and rendering, plus maps, captions, interactivity, multimedia and upgrades — with an umbrella best-practices skill when you cannot tell which fits.
4.2K stars
34 HTML Slide Templates and a Manual for Agents
Not a skill but a template library plus an operating manual for agents. Thirty-four templates hold 358 slides, each carrying twelve metadata fields: mood, occasion, tone, formality, density, and what it suits.
4.1K stars
A Goal Loop Plus 28 Marketing Skills
Two shapes: a goal loop that runs locally, and a plugin holding 28 marketing skills across Google Ads, Meta Ads and SEO, with SEO accounting for twenty of them.
3.3K stars
Claude Design, Repackaged for a Local Agent
The implementation is plain Markdown plus a few JSX scaffolds — no build step, no runtime. It detects which harness it is running in, reads the matching tool mapping, and loads only the built-in skills a task needs.
3.1K stars
AI-Writing Detection With an Engine That Scores
More than a rules file: a zero-dependency detection engine that runs in Node or a browser. The word list holds 112 entries across three tiers, alongside 61 pattern categories.
2.8K stars
Supabase's Two Skills
There are only two: one for Supabase itself, one for Postgres best practice. The latter's reference directory holds 34 documents, grouped into eight impact classes with priorities marked.
2.5K stars
A Paper-Search MCP That Says What It Cannot Download
A Python MCP server: one search tool fans out across sources and deduplicates, another handles downloads with fallbacks. A capability matrix in the README marks, from real testing, which sources support download.
2.4K stars
A Short Exercise After the Architectural Work
Once a piece of architectural work lands, it asks whether you want a ten-to-fifteen-minute exercise: predict then observe, generate then compare, trace an execution path, debug it, teach it back, or a retrieval warm-up.
2.3K stars
Driving Your Local Coding Agent From Lark
The npm package is lark-channel-bridge, not the repository name. It connects the claude or codex you are already logged into on this machine to Lark, so you issue instructions from chat.
2.1K stars
Cutting YouTube Video by Meaning, Not by Clock
It downloads at up to 1080p and chapters the video from its captions by meaning rather than fixed duration, aiming at two-to-five-minute chapters; you pick the chapters before anything is cut or translated.
2.1K stars
Deep Research That Starts From an Outline
Five commands: generate the outline, add items, add fields, investigate in parallel, produce the report. Intermediate results are JSON, and the report is a separate final step, not written as it goes.
1.9K stars
62 Skills: 50 From Books, 12 That Orchestrate
Fifty distil methodologies from bestselling books, each labelled with its source title and author; twelve meta-skills do the orchestrating, covering create, improve and grow across business, website and app.
1.9K stars
An AI News Radar Whose Reader Skill Is ai-radar
The repo actually holds two skills: the consumer-side ai-radar installs without forking, while the maintainer-side one evaluates sources and tends the crawler. The one you install is the former.
1.6K stars
Human-Voice Editing That Only Subtracts
It does not aim to improve the writing; it clears what trips the reader and leaves the rest alone. Human voice breaks into three parts: a speaker standing somewhere, judgement paid for with the body, handwork left visible.
985 stars
Deep Research in Four Depths, Up to 45 Minutes
Quick runs three phases in two to five minutes, standard six phases in five to ten, deep eight phases in ten to twenty, ultra-deep twenty to forty-five. Standard is the default, not the longest one.
969 stars
28 Mental Models, Zero Proven to Help
The entry point is a routing meta-skill, with another for combining models. The theory is credited by name to Munger, Meadows, Kahneman, Goldratt, TRIZ and OODA.
933 stars
Turning a Vague Ask Into One /goal Command
It does not execute the goal itself; it emits a Codex /goal command you copy straight out, carrying seven fixed fields: outcome, verification, constraints, boundaries, iteration policy, stop-when and pause-if.
846 stars
SwarmVault: A Knowledge Graph With Sourced Edges
One command to start: initialise, ingest local files or a public repository, compile the wiki and graph, open the local viewer. It is more than a CLI — there is an MCP server and a desktop app on three platforms.
644 stars
Cross-Platform E-commerce Skills, Count in Doubt
Plain markdown, no keys, no configuration, covering Amazon, Shopify, eBay, Etsy, TikTok Shop and Walmart. Note the count: the repo itself states 157, 142 and 125 in three different places.
595 stars
E-commerce Visual Copy Behind Two Approval Gates
Not one-click: it produces a visual plan and stops for your approval, then a storyboard script and stops again, and only after that writes the on-image copy and generation prompts.
547 stars
Turning a Month of Uploads Into an EPUB
Four steps: pull the channel's recent videos, fetch transcripts, have Claude rewrite them as magazine-style articles, package as EPUB. It is a local project you clone, configure with a .env and run with main.py.
527 stars
103 Methodology Skills in One Fixed Structure
103 skills across 16 categories, with 490 reference files. Every SKILL.md keeps the same eight sections: when to use, when not to, required inputs, the framework, workflow, failure patterns, output format, references.
518 stars
A Style Fitting Room That Picks the Direction First
It generates four mutually exclusive directions as real mini mockups; you click one and a local callback returns the choice to the workflow. Three dials adapt by task type: layout variance, motion strength, density.
458 stars
A Design System That Limits the AI's Freedom
The core claim: taste written as an operating manual. The AI invents no layouts and picks from sixteen, invents no colours and uses the brand palette, writes no styles and chooses from fifty-one components.
447 stars
Cover-Prompt Generation, Not Cover Images
A pure rules pack: what it produces is Chinese prompt text. It calls no image model itself and needs no key — generating the picture is on you. The frame is fixed at 3:4 portrait.
422 stars
n8n's Own Skills, Useless Without an Instance
Thirteen capability skills plus a meta-skill that routes between them. The prerequisite is hard: an n8n instance with instance-level MCP enabled, version 2.2.0 or newer, or none of it does anything.
407 stars
Orange-Line Illustration: One Accent, Tiny Figures
The single accent colour is fixed at #F97316 on pure white, no shadow, no gradient. Figures must be tiny — over a tenth of the frame is too large, and the author warns that models love to inflate people.
385 stars
31 Data-Analytics Skills, Scripts Included
Organised into six groups — data quality, documentation, analysis, visual storytelling, stakeholder communication, process — and what ships in the repo is runnable Python, not just prompts.
345 stars
A Video Pipeline: Three Render Tiers, Four Gates
A CLI plus a skill: emit an intermediate representation, scaffold the project, fill the timeline, then render at preview, standard or final. It defines shippable as four conditions holding at once.
309 stars
Bloom: Chasing the One-to-One Tutoring Effect
The basis is the 1984 two-sigma problem — students given one-to-one tutoring scored two standard deviations above a conventional class, into the top two percent. Note this is not Bloom's taxonomy.
222 stars
An AI PRD Skeleton of Eleven Mandatory Chapters
The output skeleton is fixed at eleven chapters, from product overview and navigation through to development priority and developer handover. Even the condensed mode writes all of them; skipping is not allowed.
203 stars
Keel: Keep Architecture's Load-Bearing Spine Small
An architecture design and governance protocol: name the few load-bearing objects, dependency direction and single source of truth, then grade what must stay stable and what should remain cheap to rewrite.
114 stars
Sixty Growth Skills for Going Global
Two layers: methodology playbooks and execution skills, the latter behind a single routing entry. The README indexes twenty-nine of them; the repository holds sixty skill directories.
66 stars
Resume Layout With the Constraints Written as Numbers
Twelve templates split eight practical and four expressive. The weight sits not in tidying content but in a set of quantified layout gates and a scripted PDF check.
41 stars
Devil's Advocate: Only Challenge, No Comfort
The whole repository is three files, and the core is a rule set that inverts the model's default agreeableness. The claim is blunt: the danger is not that AI errs, but that it ships a bad idea with equal enthusiasm.
28 stars
The Delivery Chain From Local Project to Public URL
Seven fixed steps: check CLI logins, build locally, push to GitHub, connect and deploy on Vercel, add the domain, write DNS at Cloudflare, verify, and curl for HTTPS at the end.
17 stars
Company Research That Starts From the Decision
The minimum input is two things: the subject and the decision at hand — vendor approval, renewal, competitive work, partnership, an investment screen, or a risk alert. The branch it takes depends on which.
1 star
A Weekly Review That Cross-Checks Talk Against Calendar
Evidence is ordered by availability: your spoken reflection first, then the calendar, then notes, then the long-running responsibilities. The window is stated explicitly as Monday 00:00 to the following Monday.