2. Prompting

Write prompts that reliably steer the model. Take unstructured text and produce a structured summary — first vague, then explicit. Move the same instructions from user turn to system prompt and observe the difference.

Canonical Techniques

Learn the canonical techniques by name — you’ll use all of them:

• Zero-shot — task description only, no examples. Start here; add complexity only when it fails.
• Few-shot — include 2–5 examples in the prompt. Example selection matters more than example count.
• Chain-of-thought (CoT) — ask the model to reason step by step before answering. Dramatically improves accuracy on math, logic, and multi-step problems. “Let’s think step by step” is the simplest form; structured CoT with explicit reasoning fields is more reliable.
• ReAct (Reasoning + Acting) — interleave thinking and tool use. The model reasons about what to do, acts, observes the result, then reasons again. This is the pattern behind most agent loops (Module 7).
• Self-consistency — run the same prompt multiple times with temperature > 0 and take the majority answer. Expensive but effective for high-stakes decisions.
• Prompt chaining — break complex tasks into a sequence of simpler prompts, each feeding into the next. More reliable than one mega-prompt.

Failure Modes

Know the failure modes: hallucination (confident wrong answers), refusals, and output variance at non-zero temperature.

!!! warning “The 9/10 trap” A prompt that works 9 out of 10 times is not a good prompt.

Resources

OpenAI prompt engineering · Anthropic prompt engineering · DAIR.AI Prompt Engineering Guide · ReAct paper · Chain-of-Thought paper