Rebuilt my ICM setup again, mostly because of Opus 5. This time the rules do not sit in documents. Wrote here twice before about this. It started with Opus 5. That one was by far the worst, it barely did what I asked and it ignored my own rule files completely. The other 5 models changed too, but Opus 5 is where it fell apart for me. First post I asked if others saw the same. Second post I said I had it fixed by rewriting all my rules in calmer language. That second post was too early. It ran well for a few sessions and then it was skipping rules again that it had read back to me two messages before. So I stopped rewriting rules and went looking for why it happens at all. What I learned is that a rule in a document is not a setting. It comes in like a message, and the model decides for itself whether it applies to the job in front of it. Getting the job done usually wins. And there is a limit to how much of it sticks, so every rule I added made the other ones weaker. That is why the same mistake kept coming back even though I had written a rule against it. On top of that, the one script I had that was supposed to block anything before my start steps were done turned out to have never run at all. After that I took the whole setup apart and sorted every rule by one question. Can a script check this or not. If a script can check it, it is not a sentence anymore, it is a script that blocks. If it only matters for one kind of work, it moved to its own folder and only comes in when I open that kind of file, so my script rules show up when I open a script and stay away the rest of the day. If it is a way of working I only need now and then, it became a skill that sits there as a name and one line of text until the task actually asks for it. Nothing was deleted, everything moved. The reading changed too. My big list files are not read from top to bottom anymore, they get searched. Two or three words, and only the matching line comes back. And every place where a document pointed at another document got rewritten. It used to say see file X. These models read that as a note about where something lives, not as go and read it. The older models filled that in by themselves, the new ones do not. Now it says what to do first and why.

Yes. I’m being forced to the dark side. I’m looking to push all of our learnings we have here into the framework of what’s possible inside Enterprise security. My first step is to add this post to the community and see if you will drop wisdom, repos, and guidance in the comments. Thank you kindly.

Ran a four-stage audit pipeline (my first real ICM build), assuming each stage got a fresh window. Nope. After the full run: 302.8k tokens — 281.3k of it transcript. Everything I engineered (CLAUDE.md, contracts, skills) was 25k. My architecture: 8% of the window. The chatter: 92%. One /clear: down to -282k. Messages: 137 tokens. It picked the run back up from the stage output folders without asking me anything, because my CLAUDE.md tells a fresh session to orient from artifacts, never from the conversation. (Protocol block in the comments if you want it.) So the files hold the state, but to run /clear, I have to do it between the stages? How are y'all automating the clear between stages? By hand at every gate? A wrapper running each stage as its own `claude -p`? Something else? Or does it not matter?

@Ry Mac did a GREAT job simplifying things down. Recommend taking a look. Has his own group too! https://skool.com/buildmarketclose/about

Well, welcome to one of my very few posts I have out here on this community. I've been working on a lot of different things over the last, I don't know, since last year when I really put the foot on the gas building out different accounting software. I got caught in the cycle of build, tear down, build, tear down, build, tear down, and never fully launching. So, here's my win for the last week: I was able to push my website into production. I do have a customer-facing website, but the big deal was that I was able to use a number of different coding tools to connect my CRM, build out a content creation workflow, build out the branding, and also harden my AI harness. Now I'm able to go out and not worry about prompting injections to hit my OS and the different things that might be out there when people are targeting AI agents that do things like my harness will do—going out, finding leads, improving lead identification, and the different signals for what the potential client and lead might want. The win was that I didn't do it at my desk; I did it from my phone while I was enjoying my life. That was the most important thing for me as an entrepreneur when I first started this side gig cycle of trial and error back in 2008, a few years before my first son was born. The hardest part was trying to get to a point where I could not feel guilty for working. Right now, I can do that. "Prompt and walk away" is not what these AI vampire bros want you to do right now, right? I think everyone is saying, "Hey, there's opportunity cost; we need to step on the gas. AI is cheap, computing is cheap, etc.But for me, freedom is everything. Now, I'm actively further than I've ever been—embarrassingly further than I've ever been—in all of my businesses. Reaching out has always been the bottleneck for me, so now I'm actively getting leads and actively reaching out. So the next time I post, I hope to post some deals that I've closed. I hope to share with you some of the contracts, redacted, just so that you can understand what deals are out there and how you guys can do the same for yourselves.