Chipp Walters

@grok Why is it impossible to get x.ai support to respond?

Attempt #2: Skills I have a lot of experience writing AI skills, so my next thought was: forget the giant application. Break the entire video pipeline into a collection of skills. And initially, that worked pretty well. Until it didn't. The AI would skip a step. Or forget to do something. Or decide that two steps could probably be combined. Or creatively reinterpret a workflow that I absolutely did not want creatively reinterpreted. That's when I realized that skills weren't really the answer either. What I actually needed were agents. More specifically, I needed a bunch of relatively simple agents doing very specific jobs, using very small skills. And wherever I could replace an AI skill with actual code, I did. Because code doesn't wake up one morning and decide Step 7 is probably optional. Then, sitting above all of those agents, I needed an orchestrator. And that's where this whole thing finally started making sense. The orchestrator manages the agents. The agents manage their little pieces of the workflow. Code enforces as much of the structure as possible. And the AI gets to concentrate on the things AI is actually good at. Most importantly, I don't have to babysit it. There are specific human gates built into the process where it stops and needs something from me. Those gates are predetermined. The AI doesn't just randomly decide halfway through a three-hour process that it would really like to know whether I prefer Option A or Option B. That's important because some of these processes take a LONG time. I don't want to come back two hours later and discover the whole thing stopped five minutes after I left because it was waiting for me to type "yes." So the system knows ahead of time where I need to be involved, and otherwise it keeps moving. I'm currently putting all of this through its paces by creating a very short video. I'll share that when it's finished. But in the meantime, I wanted to share the overall thinking behind this thing and where I've ended up so far. The document below shows the current plan, including the agents, steps, gates and workflow pipelines. Some of it gets pretty complicated, but I think the diagrams make most of it reasonably easy to follow. And here's probably my favorite part of the whole thing: There is no application anymore. My entire interface is a command line. That's it. I type something into the command line or drag an image into it. The system does its work and generates Markdown documents. I open those in my Markdown reader, scroll through them, look at the images and information, and react to whatever needs my attention. That's basically the entire user interface. No giant dashboard. No complicated project-management UI. No 47 buttons. No trying to remember which screen I'm supposed to be on. Just a command line going in and Markdown coming out. And strangely enough, after spending weeks trying to build the perfect interface for this incredibly complicated process, getting rid of the interface almost completely has made the whole thing dramatically easier to understand. More importantly, it's almost impossible for me to get lost. I can always see what we've done, where we are, and what happens next. Anyway, that's where I am with it right now. I hope some of you find this interesting. And once I get the first little video all the way through the machine, I'll share that too. widgetgadget.com/cw1/trash/S…