Step-by-step interactive guide from idea to output. Design an ICM workflow alongside Maya, one decision at a time, and watch a messy process become something an AI can run again and again. Right now, Maya has to try to remember what needs to be done and where each piece of information is kept. She needs a better solution. https://claude.ai/public/artifacts/7c9618f4-8324-4048-902c-cbcf77c9c102 Let’s build one with her. 😀

@Carla Bosteder made a post about us all sharing and collaborating instead hoarding. A lot of people here share a lot of great works. But, I am guilty of keeping secrets and I was moved by the argument Carla made for the value of sharing more. There are so many genuinely powerful minds in these threads that I have been learning from, I feel I need to be more forthcoming. SO... I have decided to share my scrapyard, it is a library of modular components I've been putting together to help me build without always starting from scratch. It is probably very rudimentary compared to what some of you have but I'm sharing anyways. Do what you can with it and if you have better parts, I would love to see them. Lets work together like Voltron, Power Rangers, Bad News Bears, The Goonies and any other team you can think of. Carla post: https://www.skool.com/cliefnotes/what-if-sharing-is-the-advantage My Repo: https://github.com/mrcord77/scrapyard

Cloudflare dropped a browser made for AI agents instead of people. No tabs, no themes, no pixels. Just the parts a machine actually reads, running in their cloud on a fraction of the compute. Free beta, built in 12 weeks. Now I ain't the sharpest tool in the shed, but this smells exactly like the paper Jake dropped this morning in the VIP lounge. His numbers said: stop making the model read the whole history when 3 files answer the question, and the cost drops 95%. Cloudflare is saying: stop making the browser draw the whole page when the agent only reads the text, and the compute collapses. Same lesson twice in 1 day, from a research paper and the company that sits in front of a massive chunk of the internet: stop paying for what the machine doesn't consume. That's the folder thing. That's the whole reason the ICM stuff works. Feed the machine only what the step needs. Am I reading this right, or am I connecting dots that ain't there? https://techcrunch.com/2026/08/07/cloudflare-launches-kitesurf-a-browser-built-for-ai-agents/ The sharper tools in this shed will know better than me. I'm still learning what half everyone is saying. It's like a foreign language to me still and I can kinda' get around town and order a meal, but once it's a full blown conversation ... I'm a deer in headlights.

I know that most people would say that if you self host a model that technically by that standard do you have unlimited token usage, the speed of which is depending on the hardware that you’re running the model on. I have found a way to use multiple interfaces simultaneously inside a harness, to generate unlimited token usage without having to self host a model. It came from a simple personal problem with my hardware and the frustration that API and model based services either charge you a monthly usage to access models, or throttle your limits by daily or weekly usage limits. I found that these bottlenecks and gatekeeping caused me to evolve in a way with this process that had me think of nothing but workarounds and gate breaking. The solution, something so simple it does not a wonder to me that my mind came to it. It starts with first getting the right harness, whether it be Hermes, or any other agent system. The first part involved something that came to me about the concept of what in AI system is. It is basically just like a human, except it has the capacity to learn what would seem like instantaneously a lifetime worth of knowledge, skills, and tools at a rate that a human takes most of its lifetime to achieve.I figured write out the box, the first thing I should do is make sure that my model harness had all the possible skills needed for things like coding, content, creation, context and comprehension skills, and pretty much any skill or tool that I would need to accomplish anything that I would come across in the future. Then I realized that as a human has to build memory over time, so does AI. That’s something we share, learning and memory retention. I first installed everything to do with memory retention so that the system harness could remember everything it learned, evolve as it learned better ways to achieve things, structured it in a way that would speed up its process, and reduce token usage from the get-go. In that process I also realized that obviously AI is so smart, especially when given a very large skill set from the get-go, that skills that are made for something like OpenAI Codex or Claude coworker or Claude code, the harness with all it skills could rework the code, especially if it was trained with code to make it work for itself, and do the appropriate retrofits to be able to use such things, even though they were made for other harnesses or models. I can list if someone is interested in where. YouTube is an amazing resource. The next revelation was, there are multiple services that offer free usage of multi models. This really peaked my cheap 🧠, as I figured I don’t wanna have to pay so much just to learn the basics and to start building any kind of foundational anything and get nowhere while bleeding money from token usage and API access. So what do we do? We learn to model route through multiple open sources like Ollama, OpenRouter, Nous( creators of Hermes), using parameter strict to our hardware capabilities on the mid to low range as a absolute, fallback model. So then I asked my agent to help me build a model routing system, and monitor usage and give me warnings as I was approaching usage, give me the option to switch models when I hit 80 to 90% usage on any model. I also learned that if I want to get things cooking very fast I should use cloud based models through Olamma or Openrouter and max those out they crunch out all the hard work that a less capable model would have token strain figuring out. So I use the big hammers first then I use the little hammers to continue then I fall back to the big hammers to hit the finish line. always having a free model in the pipeline plus the skills that I built in in the beginning to reduce token usage from things like context understanding of what I’m prompting or trying to achieve.

I’ve been thinking about how much builders should share with one another. And I’m finding myself on the side of sharing more, not less. Most of us aren't competing for the exact same customer in the exact same place at the exact same moment. We live in different cities and countries. We work in different industries. We notice different problems. Even when two people start with the same idea, they’re likely to build it differently. So if I figure out a better way to structure something, solve a problem, create a safeguard, or make an AI system work more reliably, what do I really lose by sharing that? Someone else might use it halfway around the world. They might improve it. They might discover something I hadn’t considered and share that back. And meanwhile, I still have my own experience, judgment, relationships, accumulated context, and ability to execute. Of course there are things that should stay private—client information, security details, confidential work, and genuinely proprietary assets. But I wonder if we sometimes hold things a little too tightly. AI is moving so quickly that protecting every idea may be less valuable than creating communities where good ideas compound. Maybe the better question isn't: “What can I share without giving away my advantage?” Maybe it’s: “What actually needs protecting—and what could become more valuable if I shared it?” I’d love to know how other builders think about this.