Why polute LLM context with large AGENTS.md if the instructions could live in the commands file?

[ahumandev]

I understand that the LLM's limited context should be carefully crafted. That is the purpose of OpenSpec.

I love the manual separation between "proposal" and "apply" commands. Because it give me a chance to evaluate whether me and the AI or on the same page.

1. However, I find it strange the OpenSpec does not use these manual commands to its advantage. Why not, setup a different context for each command which is focused on the current purpose of the command? Would that not be more efficient for a limited LLM?

2. I do not understand why the command files are tiny (23-28 lines), but the AGENTS.md file large (456 lines). Should this not be the other way around?

3. Instead of having a complex explanation trying to inform the LLM how the workflow works and what should happen at each step, why not instruct the LLM, when done with your step, call openspec next ... to get the next instruction. Then let the openspec app deal with what should be loaded from which file, instead of expecting the LLM to make multiple tool calls just to figure out what should be read. Or am I misunderstanding something?