TLDR: Brownfield development has its challenges. If you look at my previous post on specs, I mentioned that specs are very important, that you need to abstract these specs away from the code. For Brownfield projects that don’t have specs, you need to have a way to reverse engineer the implementation back to the specifications. This has been by far the largest source of improvement for agentic coding. Reverse Engineer Your Specs

Over the last month, I heard about this concept called Ralph, after Ralph Wiggum in The Simpsons. And if you don’t know, there’s this fellow named Geoffery Huntley, the person that conceived of this idea. here is the github on how to Ralph although it was not implemented by Geoffery Huntley. Here is a prompt to Reverse Engineer Your Specs

The basic idea is you run claude code or whatever harness you currently use in a loop with a new, fresh context each time you run it.

But that isn’t all that I found out.

There was more to Ralph than just the loop. You needed the artifacts to perpetuate the loop. You needed ways to generate the implementation that the loop was going to implement.

It needed and thrived off of specifications generated by a human or the help of a human, at least driving the desired result that the AI is trying to achieve.

Then it just clicked with me, I was just thinking about it for a while, and then it clicked to me that it made sense that specifications would drive artificial intelligence. Because after all, that’s how the AI gets trained, isn’t it? Off of specifications, of how well it performs a task during it’s data annotation sessions in Kenya, etc.

In my last article, I talked about how specifications were important. And I really think specifications is what’s really really important. The new way to code is through the change, additions, retractions, and mesh of specifications.

Even Peter Steinberger says that specs are the most important thing. If you don’t know who Peter Steinberger is, here is his Github. The guy is pushing 100 - 300 commits a day, +500 on a good day. This is the guy who is known for OpenClaw.

Enter AI Scaffolding and Orchestration. Software engineering is now about how to maintain a large cluster of AI bots generating code. And these guys will do some crazy things.

Scaffolding is all about creating a deterministic system that manages indeterministic behavior and actions. Creating a proactive and reactive harness that effectively controls, monitors, evaluates, sets guardrails to these indeterministic short lived souls trying to develop code.

So if you are working on a brownfield project, I would advise you to learn how toReverse Engineer Your Specs your implementation back into specs and then update the specs as you need to and use the diff of the specs as a way to drive an implementation plan to implement those specifications.

I was thinking that generating a Prompt using Prompt Wizard to ensure that the output of the prompt reached a high enough consistancy of the original project’s specs. This is a perfect use cade for prompt optimization. To tell an LLM to revere engineer the code into specs, and evaluate it against the actual specs used to generate the project.

References: Geoffery Huntley Even Peter Steinberger says that specs are the most important thing Here is the github on how to Ralph.