Skip to main content
Code on a dark screen

January 12, 2026

The Dumbest Smart Way
to Ship Code

How a bash while loop named after a Simpsons character is letting developers ship $50K projects for $300 in API costs.

The Technique

It's called the Ralph Loop. Named after Ralph Wiggum from The Simpsons· the character known for being simple, earnest, and persistently trying until something works.

$297shipped a $50K project
5,000+GitHub stars on Ryan Carson's implementation

In mid-2025, Australian developer Geoffrey Huntley was tending to his goats in the countryside when he had a realization about AI coding tools. Every time AI-generated code errored, the entire session would stall. Manual intervention was mandatory.

What if AI could act like a human·see the results of its own failed run and automatically retry based on that feedback?

The Code

Huntley's solution was almost disappointingly simple. In its purest form, Ralph Loop is just this:

while :; do cat PROMPT.md | claude ; done

That's it. An infinite bash loop that feeds the same prompt to an AI coding agent, over and over. But here's the clever part: progress doesn't persist in the AI's context window·it lives in your files and git history.

When the context fills up, you get a fresh agent with fresh context, picking up where the last one left off. The codebase becomes the memory.

"Deterministically bad in an undeterministic world."

Each individual iteration might produce garbage. But run enough iterations with clear success criteria, and the model eventually converges on something that works. Failures are predictable, and predictable failures can be fixed by tuning the prompt.

What Makes It Work

  • 1Objective success criteria. Progress isn't measured by the AI saying 'this looks good.' It's measured by tests passing, builds succeeding, and health checks completing.
  • 2State lives in files. The codebase is the memory. Git history is the log. Context window resets don't lose progress.
  • 3Failures are expected. The system assumes most iterations will fail. That's fine. Predictable failures inform prompt improvements.
  • 4Humans set strategy, AI executes. PRDs and acceptance criteria come from humans. The AI handles implementation.
Developer working late at night

The bottleneck was never the AI.
It was the human in the loop.

The Results

$297to ship a $50K project in API costs
6 reposbuilt by a YC team while they slept
3 monthsof autonomous iteration created CURSED language
Case Study: The CURSED Language

Geoffrey Huntley ran Claude in a Ralph Loop for three months with a simple prompt: create a programming language like Go, but with all lexical keywords swapped to Gen Z slang.

The result: a fully functional compiler with interpreted and compiled modes, producing binaries for macOS, Linux, and Windows via LLVM. Completely autonomous. No human code review during execution.

Case Study: Y Combinator Hackathon

A team at a YC hackathon set up Ralph loops before going to sleep. By morning, they had 6 complete repositories to demo·each with functional code, tests, and documentation. The AI ran overnight while they rested.

The Evolution

Ryan Carson, founder of Treehouse (which taught over a million people to code), took Huntley's minimal approach and added structure. His 3-file system has earned over 5,000 GitHub stars because it gives developers a process to follow.

Carson's 3-File System

1PRD (Product Requirements Document)
The human-written specification. What the product should do, acceptance criteria, constraints.
2Generate Tasks
Guides the AI to create a step-by-step task list from the PRD. Parent tasks and subtasks.
3Process Task List
Controls agent execution. Iterative code-test-commit workflow until all tasks complete.

The key insight: you're not "vibe coding" one giant prompt. You're giving the agent testable, bite-sized tickets and letting it execute like an engineering team. The AI handles implementation. Humans handle strategy and acceptance criteria.

The big idea is this: give the agent clear acceptance criteria and let it iterate until the tests pass. Not until the AI thinks it's done.

AI doesn't need to be smart.
It needs to be persistent.

Why This Matters

The Ralph Loop reveals something important about the current moment in software development. The constraint isn't AI capability·it's human bandwidth.

Before Ralph

Every AI error required human intervention. Context resets lost progress. Developers became bottlenecks in their own workflows.

After Ralph

Errors are expected and handled automatically. Progress persists across sessions. Developers can sleep while code ships.

This isn't about replacing developers. It's about changing what developers do. Less time babysitting AI outputs. More time on architecture, strategy, and the parts that require human judgment.

What would you build if the AI could run all night?

What could you ship if you weren't the bottleneck?

Share this perspective

Donny Smith

Written by · January 12, 2026

· Founder & Executive Creative Director, Bttr.

Over the past 15 years, he has led creative teams and contributed to products used by millions of people worldwide, working with companies including Apple, Opendoor, JP Morgan, GE Aerospace, Pepsi, and Alterra Mountain Company.

LinkedIn

Sources: Geoffrey Huntley on the CURSED language · Ryan Carson's 3-File System Tutorial · Ralph Loop GitHub Implementation

More Insights

Perspectives on product design, technology, and the forces shaping how people interact with the things we build.

What Is Temperature in AI
7 min read

What Is Temperature in AI

Temperature is a setting that controls how random a language model’s output is. At each step the model scores every word it knows, turns those scores into probabilities with a softmax, and temperature is applied inside that step: it divides each score before the probabilities are worked out, which sharpens or flattens the gap between the leading word and the rest. A Hugging Face guide to text generation describes the move as making the distribution sharper by increasing the likelihood of high probability words and decreasing the likelihood of low probability ones, and notes that as temperature approaches zero the sampling becomes equal to greedy decoding, where the model takes its single highest scoring word every time. OpenAI’s API documentation puts the range between 0 and 2, says higher values like 0.8 make the output more random while lower values like 0.2 make it more focused and deterministic, and recommends altering temperature or top_p but not both. The catch most teams miss is that a low number is not a lock. Anthropic states in its glossary that even with temperature set to 0 the results will not be fully deterministic, and identical inputs might produce different outputs across API calls. The operator rule follows: run cold when the work has one right answer, run warm when you want range, and never read a warmer setting as a smarter model.

What Is a Multimodal Model
7 min read

What Is a Multimodal Model

A multimodal model is an AI system that takes in and reasons across more than one kind of data, such as text, images, audio, and video, inside a single context. It maps each input into one shared representation, so it can relate a word, a picture, and a sound as parts of the same thing. The distinction that matters is how it is built. When Google launched Gemini in December 2023, it noted that the standard approach trained separate components for each modality and then stitched them together, and said Gemini was instead pre-trained from the start on different modalities. OpenAI made the same move with GPT-4o in May 2024, a single model trained end to end across text, vision, and audio so that every input and output runs through one neural network, which is why it answers audio in about 320 milliseconds rather than the seconds of lag in a stitched voice pipeline. The operator rule follows: a native multimodal model keeps every sense in the same context, while a vision tool bolted onto a text model reasons only over what it already turned into words.

What Is Tool Calling
7 min read

What Is Tool Calling

Tool calling is how a language model acts on the world. You describe a function to the model as a name, a description, and a schema for its inputs. When a request needs it, the model does not answer in prose. It returns a structured call naming the function and its arguments, your code runs that call, and the result goes back so the model can answer from something real. Anthropic puts it plainly: tool use, also called function calling, lets a model call functions you define or that the vendor provides, and the model only decides what to run, never runs it. The names differ across vendors and the substance does not. OpenAI shipped the idea as function calling, Anthropic calls it tool use, and the Claude documentation settles the overlap in one line. This is the base layer under every agent, since each single act an agent takes is a tool call, and the Model Context Protocol standardizes the layer above it so one tool, described once, can be reached by any model that speaks the protocol.