I Built Ralphex This Weekend

A first principles piece on autonomous coding loops, budget constraints, and why persistence — in code and in life — is the only strategy that compounds.

For Anyone Building Under Constraints

I'm an independent developer navigating the job market while trying to ship dependable product for Smol Nero. The tools that give you real leverage — Claude Code, Amp, enterprise subscriptions — they cost money I don't always have. And when you're building under that kind of pressure, the gap between what the well-funded teams use and what you can actually run starts to feel like the whole ballgame.

Then I found Ryan Carson's Ralph. And it dawned on me, whats something the world definitely does not need right now? A Ralph loop for their GPT Pro accounts. With aspirations in place and a whole Saturday and Sunday to work it out — I set off. Budget pressure is real. Job market pressure is real. The ambient weight of trying to stay relevant while the industry shifts under your feet is real. They're the actual conditions under which most independent developers work.

What Ralph taught me, and what Ralphex tries to embody, is that constraints aren't the opposite of progress, they are what I believe is insight on a life well lived. Every attribute you carry forward has value. The things you learned from the last iteration that didn't go well. None of it is wasted if the loop keeps running.

What I Built

Ralphex is a fork of Ryan Carson's snarktank/ralph with one core change: the runner is swapped for GPT Codex via OpenCode OAuth. If you have ChatGPT Pro, you already have everything you need. With an M1 Max, Opencode, and a frugal personality, I was set.

bash

opencode auth login ./ralphex

While building it, I hit the same failure modes Geoffrey Huntley documents in his posts. Here's what I added:

Mandatory backpressure. The original loop will silently pass if no quality checks are configured. You wake up to commits that were never actually verified. Ralphex halts at startup if no typecheck, test, or lint config is detected in your project. The loop refuses to run rather than running and producing garbage. This is not optional.

No placeholder implementations. LLMs are trained to chase their reward function, which is compiling code. Left unchecked, they'll ship stubs and TODOs that technically pass checks but don't actually implement anything. Ralphex enforces a no-placeholder rule in the prompt. If a story can't be fully implemented, it's blocked — not done.

Context discipline. Geoffrey is explicit about this: keep the primary context lean. Use subagents for file search and summarisation. Don't let large investigations bloat the main loop. Ralphex's CODEX.md now includes standing instructions for this.

👉 github.com/SmolNero/ralphex

Built on Ryan Carson's original: github.com/snarktank/ralph. Philosophy and failure modes documented by Geoffrey Huntley.

Smol Nero is an independent software studio. Follow the work at github.com/SmolNero.