Prompt Debt

We used to have messy code. Now we have messy prompts.

The Problem

You build a complex system based on a specific prompt structure: "You are a helpful assistant. Output JSON only. Use this schema..."

2 months later, you upgrade the model (GPT-4 to GPT-5), or the provider changes the model weights silently. Suddenly, your prompts stop working. The JSON output is malformed. The tone is wrong.

Why it's Dangerous

Code is deterministic. if (a > b) is always true. Prompts are probabilistic. There is no guarantee the output will be the same twice.

Mitigation Strategies

1. Eval Frameworks: You need automated tests for your prompts. Run them against 100 test cases to measure "Drift".
2. Version Control: Prompts are code. Store them in Git, not in database columns or UI configs.
3. Defensive Parsing: Never trust the LLM output. Always wrap it in a strict parser (like Zod) that throws errors if the structure is wrong.