I wrote a post about the “Ralph Wiggum” / bash-loop pattern (while :; do …; done) and why it often falls apart when you aim it at structured work (PRDs, specs, phased deliverables). It’s great for open-ended cleanup/refactor tasks, but for step-by-step workflows you run into predictable failure modes: no verification gates, weak recovery, context pollution across steps, and expensive runaway loops.

The core argument: the fix isn’t “bigger context,” it’s isolating steps and passing explicit state/artifacts between them, with verification + guardrails.

   And it's costing them hundreds of dollars per workflow — not because the idea is bad, but because the workflow shape is.

> Ralph Wiggum is phenomenal for open-ended, exploratory work:

> “Fix all the lint errors in this codebase”

> ...

> These tasks have no defined endpoint.

"Fix all the lint errors" sure looks like one.

AI slop article