I’m a Tool Builder
I’m a tool builder. I have always been a tool builder. The last 15 years or so of my career have been away from coding as my main task. GenAI has reawakened not just the coder in me, but the tool builder.
Any time I have to solve a problem for the second or third time, I’m already working on how to build a tool, system, procedure to reduce the amount of time it takes. Back in the day I was very fond of code generators. I used them, and I built my own. I built report generators as well.
I’ve been working with “Ralph Loops” - a technique in GenAI to have a complete list of tasks to accomplish and have the ai code, test, fix, test again (until all tests pass) and keep on going down the list. “Write shippable code while you sleep” is the goal. The reality isn’t quite there. The direction is clear - it’s powerful, it’s super productive. But we are less than two months from the first blog post where Geoffrey Huntley introduced the concept he named the Ralph Wiggum loop (after a Simpson’s character). https://ghuntley.com/ralph/
X (twitter) is filled with folks jumping in with their takes and improvements. I’ve reviewed several and none solve, improve the way MY mind works. So the latent tool builder in me has awakened.
My concept: Kanban board is the interface. Human enters a story, AI does a design. Human reviews, improves, accepts/rejects the design. From the design, Code Story generation. Entire goal broken down into atomic stories with clear definitions of success. Human evaluates/improves. Then TDD (test driven development) tests are written. Human reviews the tests, offers feedback, improvements. This loop goes on until human passes the tests. Then the stories with their definitions of success, the tests to prove success has been achieved - are handed off to a Ralph Wiggum loop. It will code, run the tests, fix the code, until all tests pass.
Along the way git is used to save the state at each step. And super screwup can be rolled back. Git is part of the Ralph Loop method as it serves as part of the context memory. Each iteration of the Ralph Loop starts with a clean context so that you get better code generation and use less tokens without blowing out the entire context window.
I imagine that before I am completely done - somebody will will have come out with their version of this concept. I could just wait as LOTS of people are working on this idea. But I’m a tool builder. This has me energized. I’m building a tool and LEARNING about all the GenAI concepts to make this agentic development workflow happen.
Right now I have the first step. I have kanban board where I enter a story - move it to design. A new dir in my projects folder is created, and a design.md is generated in that dir and attached to the kanban card.
This is far away from “one shot” prompting, vibe coding that has drawn attention and ire. This isn’t “look what Claude built and I don’t even know how to code”. This is team development where I am the product owner, solutions architect and lead software engineer - and I have a team of AI coding agents cooperating with me on the design, and executing the coding portion. We cooperatively test.
