Wanted to share in case it may help some of you trying to figure out how you can take the ICM framework and deploy it at an org/dept level. I work at a company of roughly 75-100 people, we're a tech company and I was tasked with building out our internal business AI capabilities. There are a lot of options, and I'm sure we'll continue to evolve and iterate the current system, but here's what works for us now:
Problem - the ICM framework is great, controlled context/routing so a model can produce expected outputs. The course and Jakes video's used Claude Code as the primary surface, given that the system itself is agnostic, it's the structure that matters. The challenge for me was "how do I take this framework that works exceptionally well for myself in a Claude Code environment, and deploy it for my entire company on one of the Claude's UI interfaces (lets be real, sales team isn't going to download VSCode)"
Initial Solution - I tried making everything it's own skill. I quickly realized bad move. Fine if you have 5 skills, breaks down at scale. You potentially run into 1) context injection problem because all skill definitions get loaded in at session start and 2) skill definitions bleed (ie. 2 or more skills with potentially similar/overlapping definitions)
What Worked:
- Surface - Claude Cowork - it's available through the main UI, anyone from our team can access it as long as they're part of the organization
- Deployment - the ICM project gets zipped and uploaded as a plugin. A plugin has a handful of bootloaders, one of which includes "skills/". This is where your project lives, and instead of CLAUDE.md and CONTEXT.md at your project root, SKILL.md acts as your high-level routing logic. Note: plugins do have a size limit of 200 MB and 5,000 files. You can use Claude Github App to propogate any merges rather than manually uploading, but unfortunately we couldn't do this due to security concerns. My plugin project structure is attached in the image provided.
- MCP - the plugin connects to an internal MCP we stood up - currently in EC2, really anything that requires oauth or secrets is managed here. Our tasks just call the necessary MCP tool
This was the simplest deployment strategy I was able to come up with for this framework. Open to any feedback or recommendations in case any of you have better implementations. Cheers