I've been struggling to get claude to follow the ICM model, it keeps jumping to summarize the files, not reading them all despite claude.md rules stating read all the files. I get caught in the same issue over and over again, it defaults to working from context and I've had arguements back and forth always ending up at the same end point, no matter what rules or check points I build it, it's not going to follow the rules each time reliably. It tells me it understands what ICM is then acts on the understanding (which is wrong despite using the git repo to build the folders) it skips steps and gates and produces the wrong outputs. Deeply frustrating I'm now testing using Hermes and pi.dev with glm flash and deepseek flash to see if open models and simpler harnesses can be better controlled. Would love any suggestions @Jake Van Clief i'm guessing you've already solved this?

@Carla Bosteder yes. Did that, actually did it twice. Ci can't get Claude to behave at all the last two weeks or so. Lewis ignoring skills, previously defined writing processes, rules like read all the files, don't grep for summaries etc Keeps ignoring all of that. I switched to Astra last week, it's been much better behaved, but still a little too agreeable

Loving this @Jake Van Clief Randomly found you on YT, dove into the ICM repo and then realised it would be a massive upgrade to my virtual COO (built initially to run in/on Claude Code but now more model agnostic). I had already started to chunk down with a routing table to what i called 'desks' but the contract idea and localised tools/skills per workroom is very smart. I am using it to help content creation for my new brand and the ability to segment down from content into written content, long form video (scripts + hyperframe editing), short form editing, is already slashing my token usage. I also think this is a much simpler way for a keen ai-newbie to get some big improvements in their workflow and output without being sucked down the 203 agents needed rabbithole.