I thought Chimera was dead... but Chimera v3.0 is back! Claude Skills brought it back to life, and stronger than ever!

📣 Hi All, 👋

I'm super excited about this one! 🚀Chimera v3.0 is back! And this is going to make your Claude (Desktop, Cloud and Code) x10 more powerful!

A while ago, I decided to abandon the Chimera Project I had been working on for weeks. The reason was that sometimes you don't need to reinvent the wheel. As I was working through it and adapting it to the new Claude Subagents, I realized that there were much better versions of Chimera out there tackling the same problems and providing better solutions. There was no reason for me to pursue it any further, and I thought the best decision was to leave the project and use what was already available.

BUT! Claude Code just released Skills this week, and this has brought Chimera back to life! 🤯

How? Let me explain what Skills are in my own words. 👇

As you know, all AI agents have system prompts where you can load instructions that the agent will follow every single time. They are very powerful, but they have their limitations. The problem with system prompts is that if you try to build a very complex system, you'll end up with a massive system prompt. Every single run will consume a vast amount of tokens, and your agent will quickly run out of context. This presents another problem: Agents are not able to use parts of the system prompt when they want; they have to run through the whole prompt every single time.

A few months ago, we got something great that has changed the way agents call and use tools: MCPs. With MCPs, we were able to pick specific sets of tools so that our agents could use a whole bunch of third-party applications when needed, without having to define specific, single API calls. We could just request that our agent send an email, and as long as it was connected to a Gmail MCP, for example, the agent would find the right API endpoint and execute the action autonomously. The problem with MCPs is also a context problem. Every time the agents have to use MCPs, they run through the whole set of tools for each MCP and have to understand what every single call does before they can execute it. This consumes an immense number of tokens, depending on the complexity of the MCP server/client configuration.

Later, Anthropic released SubAgents, an amazing, powerful tool, but only available via Claude Code. Although SubAgents are still an amazing tool, there are a few key differences that make Skills better:

✅ Zero configuration: Skills require no setup, unlike SubAgents which need files, prompts, and permissions.

✅ Maintained by Anthropic: Skills get professional updates, while SubAgents demand user maintenance.

✅ Immediate access: Skills are available instantly on claude.ai; SubAgents require a Claude Code installation.

✅ Consistent performance: Skills are optimized by Anthropic, whereas SubAgents' performance varies based on user prompts.

✅ No learning curve: Skills auto-activate, while SubAgents require technical knowledge.

Because of all the above, I really think that we could be at a "before and after" moment, similar to what we had when the MCP standard came out. 🗓️

Think of Skills as if Claude now has an internal RAG system built with a potentially infinite number of skills that the agent can pull on demand when needed, without overloading its context or system. 🧠

You are going to be hearing a lot more about Skills, I'm sure, so I hope this helps you understand them better.

And be ready to hear more about the new Chimera, because I think it's going to be a game-changer for AI Agent operability. 🎮

Soon I will be sharing the prompts and files with all of you.

See you around!

Didac

12 comments