The most popular AI programming configurations on GitHub are all public and free, and can be used by Claude Code and Codex. Ranked by number of stars:
Matt Pocock’s skills, 319,000 stars. The author is a TypeScript instructor, and this set is his daily toolbox: write tests first, write out specifications before writing code, and leave code review to two independent agents.
Addy Osmani 的 agent skills, 114,000 stars. Osmani is in Anthropic's Claude Code team, which breaks down the development process into one command per step, from /spec all the way to /ship.
Cloudflare's security audit, 67,000 stars. The agent first looks for vulnerabilities, and then a completely new agent is created specifically to try to overturn each discovery.
Obsidian skills, 51,000 stars, written by Obsidian CEO Steph Ango. Let the agent write well-formatted Obsidian notes, Bases, and whiteboards.
Elephant-Goldfish, 16,000 stars, listed under Google’s GitHub organization. An AI drafts the design, which is then reviewed by several new AIs who have never seen the chat history. The repository notes that it is not an officially supported product by Google.
Putting the five sets together, there is a common routine: let a new agent that has never seen the context pick up the mistakes. This is true for Cloudflare and Elephant-Goldfish, as is Matt Pocock’s dual-agent review. It is difficult to see the problem of the model in the same session. If you look at it on a clean instance, the effect will be better.
I don’t want to install it all, it’s enough to just move this idea into my own process.