Most business owners still treat AI like a vending machine. You type a prompt, you get a bland answer, you move on. But there’s a faster way, and it turns Claude into a full team of AI workers that actually take action.
I came across this walkthrough from Helena, an Agentic AI creator, and it stopped me mid-scroll. She breaks down how developers have open-sourced entire teams of AI agents on GitHub, and how you can install them into Claude for free. She kicks things off with a stat from Stanford’s annual AI report, which she says runs over 400 pages. Two findings stood out to her, and both matter for how you use AI right now.
The two big findings
First, the models are converging. A couple of years ago, the gap between the best and worst large language models was huge. According to the report she cites, that gap is now under 5%. Paid, free, or open source, they all perform well. AI has basically been commoditized, so the model you pick matters less than ever.
Second, businesses have not caught up. The tech is sprinting ahead while most owners still use AI as a simple chatbot. As Helena puts it, the real unlock is agents that take actions for you, not just spit out text.
Old way vs new way
Here’s the contrast that hit me. The old way is one prompt, one generic reply that sounds like everyone else’s. The new way is loading pre-built skills into Claude so it runs like a team. Helena shows the difference live. She runs the same research prompt with and without a skill installed. The plain Claude answer is vague and forgettable. The skill-powered answer pulls real links, real quotes, and specific customer complaints. Same model, wildly different output.
One key requirement she flags: these skills only work in the desktop version of Claude, not the web version. It’s a free download from claude.com, and you use the third tab labeled “code.”
The three free repos she recommends
Helena highlights three open-source projects she trusts. She also shares a smart vetting tip: check the stars and forks on GitHub. Lots of stars usually means a project is popular and worth your time.
🧠 1. LLM Council (Karpathy’s method)
This one is based on an approach from Andrej Karpathy, an OpenAI co-founder. The original idea uses multiple models like ChatGPT, Gemini, and Claude to debate each other and land on the best answer. This skill runs a version of that using Claude alone, with five advisors working in parallel: the contrarian, the first principles thinker, the expansionist, the outsider, and the executor.
Helena uses a fictional bakery to demo it. She types “console this” plus a real strategy question, like whether to chase new subscribers or raise prices for loyal ones. The five advisors debate, review each other, catch blind spots, and deliver a verdict with a clear recommendation. She says she uses it almost daily, like a free board of advisors.
🔎 2. Last30Days
This repo had over 46,000 stars when she filmed, and was the number one repository of the day on GitHub. It scans Reddit, X, YouTube, TikTok, Instagram, and Hacker News for real sentiment on any product, service, or topic.
Her example: “What are people saying about Zapier in the last 30 days?” A few minutes later she gets a full report with exact URLs, real praises, real complaints, and improvement ideas ranked by score. She points out how much richer this is than a normal Claude reply, which just gives generic G2-style summaries. Great for product research, competitor research, and spotting pain points before you build.
🛠️ 3. G-Stack (by Garry Tan)
This one is built by Garry Tan, CEO of Y Combinator, and had over 115,000 stars. It hands you an entire development team: CEO, founder, engineering manager, senior designer, staff engineer, tester, QA lead, and more.
Helena tests it by asking it to build a landing page for a fictional AI startup called Nexus. In three to four minutes she gets a finished site with an interactive 3D graphic that moves with the mouse. She notes this used to cost thousands for a designer plus thousands more for a developer, and take a month or two. Now it’s one prompt. Once loaded, you also get slash commands like office hours, design review, and debugging, so you never re-prompt from scratch.
How to get started
Based on her walkthrough, here’s the simple path:
- Download the desktop version of Claude and log in. Skills won’t run on web.
- Open the “code” tab.
- Start a new session and paste the install command for the repo you want. She provides these in her resource link.
- Approve the install when Claude asks. You may need to click allow a few times.
- Confirm it loaded. Type the slash command and look for the new skill turning blue or appearing in your list.
- Run it with a plain-English prompt, like “console this” for LLM Council or “/last30days” plus your topic.
A couple of things to keep in mind. These are community projects, so vet them by stars and forks first, and read the MD file before trusting a skill with real business decisions. And since the desktop app runs code on your machine, only install repos you actually trust.
What I love here is how it flips the whole experience. You stop being an average AI user and start running small teams that research, debate, and build for you. That’s the jump from chatbot to agent, and it’s free to try today.
Want the exact install commands and prompts? Check out Helena’s full video for the step-by-step demos and see each skill run in real time.