@chase_ai_: Everyone's hyping the Ralph Wiggum plugin wrong. The original Ralph Loop isn't just "repeat until done." It's about fresh context windows. Here's what most people miss: LLMs get dumber as the context fills up. Past ~100k tokens, effectiveness drops hard. The REAL Ralph Loop starts a brand new session for every task iteration. Fresh context = smart Claude. Stale context = dumb Claude. How it actually works: Create a PRD with discrete tasks Ralph starts a NEW Claude Code session It reads the PRD, finds the first incomplete task Codes until done (or fails) Updates a progress.md file with what worked/failed Starts ANOTHER new session Next iteration reads progress.md → knows what NOT to try again The power isn't repetition. The power is fresh tokens + learned context from previous attempts. The Claude Code plugin? It just blocks the exit and keeps going in the SAME session. No fresh context. Misses the whole point. That's why the creator has been posting corrections everywhere. Real Ralph = new session each iteration. Plugin Ralph = same session until auto-compact. One keeps you smart. One lets you drift into the dumb zone
