The Wisdom Wall

“I generally wouldn't go over 50%, meaning like the anthropic model Opus 4.5 has a 200,000 token context limit. The moment, in my opinion, you've got over a 100,000 tokens, meaning you're using the same session. It starts to sort of deteriorate. That's when you have people, Greg, who say, oh, like I started off good,…”

“I think the moat for the next couple of years, cause, and I, I'm willing to argue this is design. Design and user experience becomes the moat. Um, cause at the end of the day, like if you can make the experience much richer, much better, and you can actually solve the problem, um, you know, that's, that's where you…”

“And believe me, when you give the agent the floor to give assumptions, most of the time it's going to get it wrong, but not only is it going to get it wrong, it's going to burn a lot of money.”

“when you and I are trying to use AI to build something meaningful, I hundred percent, hundred percent, hundred percent stand in the fact that the human still needs to be in the loop. AI can replicate sauce. It can't create sauce.”

“the only place a loop makes sense is in a very confined, constrained process with a very fixed feedback loop, a very defined feedback loop, and that's in code review.”

“I don't download skills because your agent needs the context of a successful run, which you then turn to skills, right? Um, and this is the big thing I see. You see skills marketplaces. You see download this and that. First of all, it's a easy way to attack somebody. So I would be very, very careful with downloading…”

“So there's like this early area of investment that you have to make that sucks that nobody will tell you, especially agent harnesses company because they wouldn't raise as much money if they did. But like this, maybe I would give it two weeks. Because it took me two weeks.”

“don't over obsess. Obsess on MCP skills, et cetera, et cetera. I'm not saying don't get into these. I'm not saying don't read about them. I'm not saying don't use them, but I can almost guarantee you these things are not the reason why your product isn't working, right? Most of the time it's your plan sucks, right?”

“If you're wondering why we don't have an Ironman level Jarvis assistant is because combining these tools, making it work with the LLM is one thing. But then stacking these tools on top of each other, making it cohesive, making it work together is a nightmare itself.”

“I realized, Greg, that a lot of people are, um, are led down by the AI tools hallucinating or giving them the wrong thing, and that's because most people are terrible product managers.”

“I truly believe if you're trying to build something, um, that you, uh, that you want users to use, you're going to have to get your hands dirty.”

“anytime I push over 1000 lines of code, One K lines of code. Like if the, if the code that it has to review is more than 1000 lines, I can almost never get a five out of five because it's too much code for the agent to fully review and contextualize and understand.”

“to me, the best way to create a skill is to work with it in your specific workflow. Once you have a successful run, tell it, okay, review what you just did. This is the skill you need to create.”

“we should treat models and these agents like very new employees versus like these black magic boxes that like know everything. Right? They know everything because they've been trained on a lot of data, but they don't know your workflow, your steps, right?”

“most of the time, you're sort of Allowing the AI to have free reign over certain decisions, which I think, uh, will lead you with a finished product that you're not excited about.”

“if you invest the time in the planning stage, I hundred percent believe you'll save a lot more money, and this will help you clear up a lot of ideas.”

“And this is why a lot of people who were fighting with Claude code all these months are really, really good at using it now because they spent the time building without using these crazy automation loops.”

“If you have a terrible plan, if you have a terrible PRD, This doesn't matter. You're just donating money to Anthropic, and I wish you the best of luck if that's what you want to do. But if you want to make sure that your tokens are not wasted, you're going to invest in a good prd.md file or a good plan file.”

“if we were to sit here eye to eye and you haven't built anything, deployed anything, there isn't a URL that I myself or Greg can click on that you've built, you have no business using Ralph. You literally have no business using Ralph. I would first get good at prompting and building something using a plan Whether it's…”

“I think a little bit of audacity, a little bit of thought and care, and a little bit of taste goes a long way nowadays, um, and more than the models getting better, because it's going to get easier, it's going to get better, it's going to get faster, but unfortunately, if you don't change, then it doesn't all matter.”

“And you've put in the groundwork to keep testing, iterating, playing, breaking, learning that when they do become really, really good, you're not starting from scratch. You've already have this knowledge. Now you're starting to build and ship fast.”

“Now the reason why it's asking me to review this cause SQL is no joke. Um, you can really destroy business writing the wrong commands”

“Cause again, you can't always assume the AI knows everything. And this is where, um, I know we just jumped straight into bolt, but if I wanted to build an actual application, I would do research first on what I need, how it would work, how it works, what the setup looks like. But here we're just trying to have fun and…”