My 5-Step Forge for Better LLM Coding for SWEs

I’ve been finding much improved results using LLMs to code by disagreeing with the core tenet of “vibe coding” — that you can “one-shot” code. You should not, in fact, try to “one-shot” code. Here’s how to break the LLM of that habit and get better code.

Professional code isn’t made in a single step, which is why I created my 5-step forge prompt for SWEs using LLMs to code. (Mathieu Odin for Unsplash+. Unsplash+ license, used with permission.)

If you’ve ever run code generated by an LLM without reading it first, only to then read the code and identify that the LLM’s whole plan and architecture approach were wrong — then this is the article for you.

This is not the article for you if you’re looking for tips on how to better “vibe code” a greenfield project from scratch, though the five-step prompt I recommend in this article will also work for that use case.

This article is for my fellow SWEs.

It’s for those of you who are working every day with an LLM and, like me, are getting frustrated when we tell the LLM, “that didn’t work,” and the LLM says, “Of course it didn’t work! There’s an obvious problem!”

This five-step prompt will prevent that problem, a significant portion of the time: 1PLAN, 2CHECK, 3CODE, 4CHECK, 5RUN.