Why Your AI Prompts Break After One Edit (and the Versioning Fix)

You spend 30 minutes crafting the perfect prompt. It works beautifully. Then you tweak one line, and the output goes sideways. Sound familiar? The problem isn't the edit. It's that you're treating prompts like scratch notes instead of versioned artifacts. The Fragility Problem Prompts are deceptively sensitive. A small wording change can shift the model's interpretation dramatically: "List the top issues" → bullet points "Describe the top issues" → paragraphs "Analyze the top issues" → a 2000-word essay Unlike code, there's no compiler to catch these shifts. The prompt "works" — it just produces something different than before. The Versioning Fix I started version-controlling my prompts the same way I version code. Here's the lightweight system: 1. One File Per Prompt prompts/ ├── code-review.md ├── bug-fix.md ├── test-generator.md └── changelog.md Each file has a header: # code-review v3 # Last working: 2026-04-01 # Changed: Added "skip style nits" constraint # Previous: v2 — too many