8 GitHub repos for AI agents: what they need, and our test of two skills

Greg Isenberg makes a popular podcast about startup ideas. In a new video he shows eight free GitHub projects for AI agents, and how he would use each one to make money. GitHub is a website where people share code. Do these tools really help? We looked at all eight and tested two of them ourselves. The Grill Me skill clearly changed how Claude asked its questions. The Free Tools skill hardly changed the advice. And three of the other six need more care than "free to download" suggests.

What the video is about

The video came out on 8 October 2026 and is about 24 minutes long. Greg Isenberg walks through eight projects that he tries "at night and on weekends". For each one, he explains what it does and one way to earn money with it. For example: make a launch video for a company before you even talk to them, so you have something to show.

He says most of them are free, and that you pay for the AI model you use with them. His picks are Brag and Agent Reach. He also shares a sum: ten clients paying $1,000 a month gives $10,000 a month. That is an example of what could happen, not proof that it works.

The eight repos in short

A repo is a project folder on GitHub. These are the eight, with what they need, from their own pages on GitHub:

• [Brag](https://github.com/latent-spaces/brag): makes a short launch video from your project. Needs Node.js 22 and FFmpeg, a video tool.
• [Agent Reach](https://github.com/Panniantong/Agent-Reach): lets an AI agent read X, Reddit, YouTube and more. For X you export your own login cookies.
• [Voice Studio](https://github.com/debpalash/VoiceStudio): clones voices and dubs videos on your own computer. Needs about 10 GB.
• [PPT Master](https://github.com/hugohe3/ppt-master): turns notes into a real PowerPoint file. Needs Python and an AI model.
• [OpenRig](https://github.com/mvschwarz/openrig): lets several coding agents work as a team. Runs on Mac and Linux, not on Windows itself.
• [Free Tools](https://github.com/coreyhaines31/marketingskills/blob/main/skills/free-tools/SKILL.md) by Corey Haines: helps you plan a free tool that brings in customers.
• [Grill Me](https://github.com/mattpocock/skills) by Matt Pocock: lets the AI question you about a plan.
• [Impeccable](https://github.com/pbakaus/impeccable): design advice for AI agents.

All eight have an open licence, so you may use them for free. Most have tens of thousands of stars on GitHub, a sign that many people saved them.

What is a skill?

Two of the eight are skills. A skill is a text file with instructions for an AI helper like Claude Code. You put it in a folder, and the AI reads it when it fits the task. You do not install a program, and nothing runs on your computer.

The grilling skill behind Grill Me is only 30 lines long. It tells the AI to question you in rounds, number every question, and give its own recommended answer each time. You can then just say "yes" to agree. That is why we picked the skills for our test: you can read every word before you use them. We did, and found no code and no hidden steps.

Our test: the same question, with and without a skill

We gave Claude Code the same question twice: once in an empty folder, and once in a folder with the skill in it. We did each one two times, so eight runs in total. Claude Code is Anthropic's AI helper for programmers; read more in our Claude Code pricing article.

• Grill Me: "Grill me on this plan before I build it: a WhatsApp bot that lets customers of small air conditioning installers book a maintenance visit. I want to charge installers 49 euros a month."
• Free Tools: "I run a small air conditioning installation company in Limburg, the Netherlands. I want more leads from my website. Should I build a free tool for that? Give me a concrete plan."

We used the Sonnet 5.5 model with the default settings. Each run took 17 to 27 seconds. In every run with a skill, Claude really used it.

Grill Me: questions that come with answers

This skill made a clear difference. With Grill Me, Claude asked 7 and 6 questions in the first round. Every question came with a recommended answer, like "Yes: use the official WhatsApp API. Never use unofficial tools for a paid product." At the end it said what the next round would be about, and it waited.

Without the skill, Claude was not bad at all. It asked about the same risks: WhatsApp costs, who sets the calendar, what happens when a customer reports a leak. But it put 10 and 8 topics on the table at once, often with several questions each, and mostly without a recommended answer.

So the skill does not make Claude smarter. It makes the questions easier to answer: you can say "yes" or "no" to each one and move on.

Free Tools: one idea, four times

Here the skill made little difference. In all four runs, with and without the skill, Claude came up with the same tool. It was a calculator on the website that tells visitors which size of airco they need and what it will roughly cost. Then a button to ask for a quote.

The details differed a bit. With the skill, Claude showed the result first and asked for contact details after. One run gave the idea a score: 29 to 30 out of 40. Without the skill, one run hid the price until you filled in a form, and the other showed the result first too.

For a question like this, Sonnet 5.5 already knows the usual advice. The skill mainly adds a checklist.

What the Free Tools skill adds

Corey Haines' skill is mostly a checklist. It asks the AI to rate an idea on eight points, from 1 to 5. For example: do people search for it, does it fit your buyers, and is it easy to keep running? 25 points or more means a strong idea.

It also lists famous examples, like HubSpot's free website grader, and warns against building too much. That is useful if you are new to this, or if you want the AI to check its own idea in a fixed way. If you already know your customers, the plain AI got to the same idea.

Before you install the other six

We did not install the six programs. "Free to download" is true, but their own pages show what you take on:

• Agent Reach uses your login cookies for X and your logged-in browser for Reddit. Its README advises using a separate account, because your account could get banned. Web pages go through outside services (Jina Reader and Exa).
• Voice Studio downloads Python, FFmpeg and AI models on first start, about 10 GB. It asks you to clone a voice only with permission.
• OpenRig changes settings files of your terminal and your AI tools, and does not run on Windows itself.
• Impeccable downloads a program the first time you run it.
• Brag and PPT Master need Node.js or Python set up first.

None of this is wrong. But it means you should try them in a test folder or a separate account, not on your main computer with your real logins.

How we checked

On 8 October 2026 we watched the video and read the description on YouTube. We read the READMEs of all eight repos on GitHub, and the full text of the two skills we used: grilling and free-tools, with its two reference files. The screenshots are our own captures, cropped only.

We ran Claude Code 8 times on the site owner's Claude plan, with the Sonnet 5.5 model and default settings. We turned off the owner's own skills, but the owner's general instructions still loaded in both versions; they include that the owner runs an airco company. That is why both versions knew about it. We asked for answers in English. The charts were made by us. None of the images was made by AI.

This was a small test with two questions and two tries each. AI answers change from one try to the next.