39. Prompt Engineering for Experts (The Real Stuff)

Beyond "You are a helpful assistant"

Expert prompting is about treating prompts as programmable interfaces. You version them, you test them, and you structure them hierarchically.

Key Techniques

• Chain of Thought (CoT): "Think step by step" is the "Hello World" of CoT. We'll learn how to force structured reasoning blocks (e.g. `...`) before the JSON output.
• Few-Shot Selection: Dynamically injecting the 3 most relevant examples into the prompt based on the user's query (using embeddings).
• Self-Consistency: Asking the model the same question 3 times with high temperature and taking the majority vote.

Where to go next

• 39.1 Hierarchical prompts: global rules → module rules → task rules
• 39.2 Prompt “APIs”: stable interfaces between app and model
• 39.3 Self-critique patterns (and when they backfire)
• 39.4 Multi-step reasoning with enforced intermediate artifacts
• 39.5 Prompt compression and distillation