System prompts fail in predictable, mechanical ways. This post separates beginner mistakes from advanced ones, with concrete fixes, code, and a verification loop for each.

Retrieval failures in RAG pipelines are not random events — they follow patterns. Here is a field guide to the seven failure modes I see most often, with concrete fixes for each.

A single model has a predictable failure profile. An ensemble of models can cancel those failures out — if you wire it correctly. Here is a practical checklist for combining outputs.

Prompt quality is not a vibe. Here is a concrete system for measuring response latency, output consistency, and task completion rates so you can compare prompts with data instead of gut feel.

You are probably paying for tokens you already paid for. Here is how prompt caching and a few upstream habits can cut your LLM API bill by 40% or more.

JSON mode still breaks in predictable ways — malformed output, wrong types, dropped fields. Here's a symptom-by-symptom guide to diagnosing and fixing each one.

A single line of hidden text in a PDF was enough to make a support bot leak internal pricing data. Here's a full walkthrough of how that happens and what actually stops it.