I Stopped Writing Prompts as Plain Text — Here's What I Do Instead

I've been doing a lot of prompt engineering lately, and I hit a wall that I think many developers will recognize.
My prompts started looking like messy config files. A system role at the top, then task instructions, then constraints, then output formatting rules, then few-shot examples. Every time I wanted to test a small change — like swapping the persona or adjusting one constraint — I was duplicating the entire prompt and carefully editing one section. It felt like editing a monolithic codebase with no modules.
So I started thinking about prompts the way we think about code: modular, composable, reusable.
What if each section of a prompt was an independent block? What if you could toggle a block on or off to A/B test its impact? What if your "output format" block could be shared across 20 different prompts?
That idea turned into a tool I've been building called Prompt Builder. It's a block-based editor where prompts are assembled from draggable, reusable components instead of written as a wall of text.
I'm curious whether other devs have felt this friction or if I'm solving a problem that only exists at a certain scale of prompt complexity. What does your prompt workflow look like?

promptengineering #ai #productivity #webdev