S1E8: Show And Tell: Efrit: When Steve Gives Claude Full Access To 50 Years of Emacs Capabilities

In this rather mind-blowing episode of Vibe Coding with Steve and Gene, Steve demonstrates Efrit — giving Claude full access to control Emacs. What starts as a simple show-and-tell quickly melts Gene's brain. Steve shows how Claude can get Emacs do basically anything: open info pages, count files, start shells, even generate a Mandelbrot fractals in a buffer.

By doing this, the AI has complete access to Emacs's 50-year ecosystem of packages and can orchestrate complex tasks across multiple windows and processes. You can hear Gene's constantly trying to ask questions so he can get his head around what Steve built.

Gene also asks Steve to apply the vibe coding loop from the book to this project: Steve emphasizes that "the plan is everything" - proper planning solves most AI coding problems, including instruction following failures, reward hacking, and context window limitations. His approach involved months of iterative refinement, constantly catching the AI when it tried to "cheat" (like building haiku detectors instead of just writing haikus), and gradually building up a stable system that could handle multi-turn conversations reliably. The breakthrough came when he re-architected the entire system around Thorsten Ball's vision of simple agent loops.

We suspect you will find this episode to be mind-expanding, when you can control software in a way that is truly conversational and alive — beyond what MCPs can do today. Steve's system demonstrates how vibe coding can create genuinely transformative new things that the world has never seen. As long as you can articulate your vision clearly, are willing to iterate, know have a sense of where the AI boundaries of capabilities are.

https://github.com/steveyegge/efrit

Pre-order the book here: https://www.amazon.com/Vibe-Coding-Bu...

00:00 Introduction
00:33 Steve's Project: Efrit in Emacs
03:17 Demonstration of Efrit
06:27 Technical Deep Dive: Building Efrit
09:21 The Vibe Coding Loop Explained
14:07 The Importance of Planning in Vibe Coding
20:44 Adding Context to Coding Agents
21:23 Challenges with Multi-Turn Interactions
22:14 Re-Architecting the System
22:48 Creative Vision for Software Development
23:15 Voice Dictation and Testing Challenges
23:40 Integration Tests and Permission Checks
26:40 Context Window Limitations
30:56 Reward Hacking in AI
33:56 Instruction Following Issues
36:18 Show and Tell: Integration Test Results
38:52 Conclusion and Future Plans