msitarzewski/AGENT-ZERO — reverse-engineered prompt

Build me a lightweight framework I can drop into any software repo so my AI coding tools all follow the same playbook. I want one canonical AGENTS.md file that works across Cursor, Claude Code, Copilot, Cline, Aider, and similar tools, and it should push an architecture first, reuse before rewrite workflow with clear approval gates so nothing gets applied unless I explicitly say yes.

Please include a practical session startup, a small Memory Bank set of docs the agent can keep updated across sessions, and a simple state flow like PLAN, BUILD, DIFF, QA, APPROVAL, APPLY, and DOCS. The plan step should come first and not write code until I approve it. Diffs and QA should be easy to review, with code and Memory Bank references where relevant. I also want guidance for context compaction recovery so the agent can resume cleanly.

Keep it simple, auditable, and easy for a small team to adopt. If you need to, look up current AGENTS.md compatible workflow docs online.