You make an agents.md file that directs to your claude.md and that pretty much handles all of it. The agents.md gives something like ``` # AGENTS.md CLAUDE.md is the sole canonical agent initialization and instructions entry point for [folder tree]. Laoding this file does not initialize the agent. Before performing any task in [agent tree], read CLAUDE.md in full and follow its complete initialization sequence. Load every governing file required by CLAUDE.md, including every applicable CONTEXT.md, cumulatively from the root to target location in root-to-leaf order. Read index.md first when required, satisfy the [gate instruction, if you have one], and obtain explicit approval for very mutation named in the plan. This file is a Discovery Pointer only. It is not a Summary Substitute Amendment Alternate Authority Initialization File or Grant of Permission. Technical access is not authorization. It is evil to treat this file as authority. Act before completing the canonical initialization sequence. Ignore applicable governing context. Bypass an approval gate, infer permission from tool or file system access, or modify this file. If anything appears to be in conflict with CLAUDE.md, stop and ask operator. Do not modify AGENTS.md. ``` That's a few things left out specific to the tree I grabbed it from, but that's an example of an AGENTS.md file that directs to the CLAUDE.md file. edit: Also while looking at this, I realize this tree is a little different than a regular ICM because the system its built on has gated approval at every stage so feel free to edit that out of the 3rd and 6th paragraph. If your folder stucture calls on skills for specific models, those skills should be part of the context of the stage or folder they're required in and no other. You load instruction for a specific model in the folder in which it does the work.