Ralph Loops: Build Dumb AI Loops That Ship — Chris Parsons, Cherrypick
TL;DR
Chris Parsons’ core move is dead simple: put the model in a loop and let it keep going — a “Ralph loop,” named after Ralph Wiggum, is basically “implement the ticket” repeated until the model notices what it missed and fixes it.
He argues modern models made loops practical only recently — Parsons says this starts really working with newer frontier models like GPT-5.12+ and Claude Opus/Sonnet 4.6+, where the agent can reliably read, act, re-read context, and decide the next step.
The big unlock is not re-running one prompt, but letting the AI pick the next most important ticket — instead of hand-orchestrating dependency graphs or parallel agents, he tells Claude to scan
doc/tickets, choose the next priority, implement it with TDD, mark it done, and commit.His newsletter workflow is the before-and-after story — he replaced a brittle n8n automation that failed “2 p.m. every Monday” with a Claude Code skill that effectively loops through the job and now produces a much stronger first draft with far less maintenance.
Parsons is using loops far beyond coding — he says he runs them “24 hours a day” for email drafts, calendar checks, newsletters, client work, a 6 a.m. morning briefing, and even a “worker loop” that advances project files inside a personal work vault.
The real bottleneck isn’t tokens, it’s deciding what humans should still do — his rule is to let AI handle reversible work like drafts and slide decks, but keep irreversible or identity-defining work like strategy, final sends, and reputation-sensitive decisions for himself.
The Breakdown
From Weird Name to Serious Workshop
Chris Parsons opens by treating the room like a live lab, not a lecture: grab your laptops, we’re building this together. He frames the moment with a revealing show of hands—lots of people already use Claude Code, Codex, or Cursor for code, and a surprisingly large number say they’re “no longer writing any code,” which he calls a huge shift compared with even six months ago.
The n8n Automation That Kept Ruining Mondays
He contrasts his old AI workflow with what he does now: a giant n8n newsletter pipeline that took a week to build, was painful to debug, and reliably broke every Monday afternoon. The key emotional beat is that it became easier to write the newsletter himself than maintain the automation that was supposed to save him from writing it.
Why Skills Beat Workflow Orchestration
Parsons says the future isn’t elaborate orchestration graphs; it’s loops with good context. He copied the n8n JSON into Claude Code, asked it to turn that into a skill, and got a simpler system that reads instructions, takes a step, re-reads, takes the next step, and eventually stops with a better first draft than the old workflow ever managed.
The Original Ralph Loop: Just Do the Same Thing Again
He traces the idea back to Jeffrey Huntley and the absurdly simple insight from “ancient AI times” — around last June — that when a model says it’s done, you should just ask it to do the same thing again. That’s the Ralph Wiggum joke: keep trying the same thing, and often the model uses the second or third pass to notice the tiny thing it forgot.
Live Demo: A Bad Pomodoro Timer Becomes the Testbed
To make the concept concrete, he opens a tiny Python Pomodoro app he “vibe coded in about three minutes” and points Claude at ticket 001, which adds a status command. The demo is delightfully messy—conference Wi-Fi dies, he tethers to 5G, Claude eventually implements the feature, adds a test unprompted, and on the second pass notices it forgot to mark the ticket done, which perfectly proves the loop’s value.
The Important Upgrade: Don’t Pick the Ticket Yourself
The real jump, he says, came from Matt PCO’s framing: stop obsessing over whether one ticket is done and instead point the loop at a pile of work. Parsons describes his failed attempt to pre-plan a giant dependency graph with multiple agents as basically recreating 1990s waterfall requirements docs; the better move is one line: “implement the next most important ticket.”
Loops Everywhere: Cron, Skills, and the Worker That Never Sleeps
From there he expands the idea far beyond code. He shows Claude Code’s /loop command running every minute or hour, imagines checks like “look in Linear for new bugs from test,” and explains that much of software development is already just backlog-review-build-PR-release-start again. He then zooms out to his own system: a 15-minute heartbeat bot, a 6 a.m. daily briefing, and a “worker loop” that advances project files, drafts replies, and prepares work without taking irreversible actions.
The Existential Turn: If Everything Is a Loop, What’s Left for You?
The workshop ends on the most human note. Parsons says these systems forced him to ask what he actually wants to do: he doesn’t want to become “the email reviewer,” even if AI can tee up 15 or 16 draft replies overnight, so he protects strategic thinking and judgment as the work he still claims. His practical rule is memorable and sharp: if an action isn’t reversible without embarrassment, the AI doesn’t get to do it.