I found 10 trending skills on GitHub that can boost your agent power a lot, and I had to test them this week :)
Not your skills. Agent Skills. Small folders with a SKILL.md inside that teach your AI agent to do one thing really well, again and again. You write it once, any agent can use it next time.
Today GitHub Trending is full of them. Top 5 are literally all skills repos. I opened each one, read the README, checked stars and use cases. Here is what I found.
Quick intro from me. I am Taqui, I build web apps with Next.js and TypeScript. I also made Wryte, a small AI blog writer for myself. So I keep an eye on what helps agents build better, not just chat better.
What is a Skill in simple words
A Skill is not a plugin. It is instructions, scripts and resources in a folder. Your agent checks if it needs that skill, loads only that folder, follows the steps.
Anthropic started it. Now Google, Vercel, Cloudflare, Meta, even Prometheus ship one.
Why does it matter for you as a Next.js dev? You stop copying the same prompt in Claude Code, Codex and Cursor. You write the skill once in SKILL.md. That is it :)
Below are 10 trending repos, with real stars, real use cases and links. Save this for later.
1. Cloudflare security-audit-skill - 66K stars, 2,085 today
What I found: This skill turns your coding agent into a security auditor. Six phases: Recon, Coverage-led hunting, Candidate validation, Structured output, Independent record verification, Target-neutral reporting.
It writes findings.json and coverage-ledger.json. Fresh agents try to disprove each finding. That is how Cloudflare built their vulnerability harness.
Use case for you: If you ship auth in Next.js, run this before you push. Even phases 1 and 2 help a lot.
Repo: cloudflare/security-audit-skill
2. Matt Pocock skills - 317K stars, 1,100 today
What I found: Matt Pocock put his real .agents folder in public. Tagline: "Skills for Real Engineers, not vibe coding."
Small, composable skills. Run /setup-matt-pocock-skills once, it asks where you keep tickets and labels.
Use case: Triage, docs, issue planning, the boring parts you repeat every project.
3. Anthropic skills - 188K stars, 390 today
What I found: The original repo. It has the skills that power Claude's docx, pdf, pptx and xlsx features. Also creative skills for art, music, design and technical skills for web testing and MCP generation.
Each skill is a folder with SKILL.md. All Apache 2.0.
Use case: Learn what a good skill looks like. Copy the template/ folder to make your own logo, doc or data skill.
Repo: anthropics/skills
4. Google skills - 23.5K stars, 443 today
What I found: Agent Skills for Google products, including Google Cloud. One command installs many: getting started, IAM, AlloyDB, Agent Platform, RAG for enterprise search.
Use case: You build on GCP or use Gemini. Instead of pasting docs into prompts, you install the skill once.
Repo: google/skills
5. Vercel agent-browser - 47K stars, 190 today
What I found: Browser automation CLI for AI agents in Rust. You open a page, get a snapshot with refs like @e1, click, snapshot again. The browser stays open.
Install: npm install -g agent-browser then agent-browser install
Use case: Let your agent browse docs, fill forms, test your Next.js app. I used similar thinking for Wryte experiments, refs are way more stable than CSS selectors :)
Repo: vercel-labs/agent-browser
6. Meta astryx - 14.7K stars, 96 today
What I found: Open source design system from Meta, agent-ready. React 19 and StyleX, 150+ accessible components, theming, dark mode, CLI. Used internally 8 years across 13,000 apps.
Components can be ejected as full source. No lock-in.
Use case: If you build design systems for agents, this shows how to ship component docs an AI can actually use, not just code.
7. Sherlock - 101K stars, 1,766 today
What I found: Hunts social accounts by username across 400+ networks. sherlock user123 writes user123.txt.
Use case: Check your own username leaks, or do OSINT if you work in security. Use it responsibly please.
Repo: sherlock-project/sherlock
8. OpenMausBot - 17.3K stars, 1,955 today
What I found: Open-source Grok Bot alternative with a VM bots can use. It has .claude/skills/ and skills/, plus computer use drivers. Idea: give the agent a real computer, not just text tools.
Use case: Automation that needs a mouse, keyboard and files. Works with Docker or Podman self-hosting.
Repo: milind-soni/OpenMausBot
9. Prometheus - 67.4K stars, 322 today
What it is: Prometheus monitoring and time series database. Collects metrics, evaluates PromQL, alerts when conditions match.
Use case: Your agent needs observability. It now ships .agents/skills so an agent can learn scrape, query and alert without you pasting docs.
Repo: prometheus/prometheus
10. Ruff and billion-context - 51K + 4.7K stars, 350 + 191 today
I put these two together because both save tokens, which is money.
- Ruff: Python linter and formatter in Rust, 10 to 100 times faster than Flake8 and Black, 900+ rules.
- billion-context: Context compression for month-long sessions, 100K window is enough, about 5 times fewer tokens.
Use case: Lint fast, keep context small, ship cheaper AI features.
Repo: astral-sh/ruff and ranxianglei/billion-context
Final thoughts
I do not know which skill will win. That is OK. The point is you can copy one today, learn, and make your agent better this week.
Are you already using a skill in Claude Code or Codex, or still copying prompts by hand? :)
If you want to start, read Anthropic skills and Google skills. Worth a look :)
I'm Taqui, full-stack dev from India building web products with Next.js and TypeScript. I post what I'm learning here and on my portfolio at taqui.in. If this helped, say hi in the comments :)
Top comments (7)
66k stars on a folder of markdown instructions, no way that's all organic. My bet is none of these survive the month and the skill that sticks is the boring one you write for your own repo's weird deploy steps. Did any of the ten survive your test week?
Bro 😂, I think you are outdated
The #1 is really an awesome skill and you are telling organic
The Dev's those have dropped star's on the skill repos are not mad... Try to understand
I really don't want to copy propmt or skills also. Because I building the stuff on my own intention.
The Cloudflare one is the only entry here I'd actually run before pushing — a fresh agent trying to disprove each finding is the whole idea, everything else is just a checklist with extra steps. My bet is the same as ivannovazzi's though: the skill that sticks is the boring one you write for your own repo's weird deploy steps, not a 66k-star folder. Did any of the ten survive your test week, or is the list already out of date?
A skill list becomes more useful when every entry has a boundary as well as a capability. I look for its required inputs, side effects, verification step, and explicit stop condition before installing it. That small contract prevents a collection of impressive prompts from becoming an unreviewable source of authority.
The Cloudflare security-audit-skill section stood out to me, especially the idea that fresh agents try to disprove each finding and that the results go into findings.json and coverage-ledger.json. That second-agent check seems like the key part — otherwise an audit skill could just produce a long list of plausible-looking issues. How do you handle it when the fresh agent disagrees with the original finding: is the finding dropped, downgraded, or kept with a note about the disagreement? I am also curious whether phases 1 and 2 alone are enough to catch the common Next.js auth mistakes, or whether most of the value only shows up once the coverage-led hunting in the later phases runs.
Awesome post
I need to learn more about SKILLS, I only recently realized how important it is for each agent to have its own skills. 😀