chaseai-yt/claudex-loop ? reverse-engineered prompt

Build me a Claude Code skill and plugin that helps me harden a software plan before any code gets written.

I want the flow to be simple for a normal user, first the agent does a quick recon pass, then it asks me only the important questions one at a time, then it locks the plan into a PLAN.md file and keeps a full review log. After that, a second model should review the plan in a read only way, point out flaws, and force revisions until the plan is solid or it hits a clear round limit. If I choose to build from the plan, one model should write the code and the other model should inspect the final diff and compare it against the plan.

Make it feel usable from Claude Code with easy commands or skills, and include a plugin setup plus a manual copy option. Keep the process clear, opinionated, and safe, with the model that writes never grading its own work. If you need to look up current docs online, do that.