Phase 2 — Working with LLMs · Lesson 24 · 20 XP
Prompting as engineering
Good prompts are specific: clear instructions, a defined output format, and — often more effective than another paragraph of instructions — one or two concrete examples (few-shot). An example shows the model exactly what "good" looks like instead of describing it abstractly.
Common failure modes: ambiguous instructions the model has to guess at, a system prompt that quietly contradicts a later user instruction, and stuffing in unrelated context that dilutes what actually matters. Treat prompts like source code — version them, diff changes, and know what changed when behavior changes.
Exercise
Take a vague prompt and identify exactly what's ambiguous about it. Rewrite it three ways — more specific instructions, one added example, and a constrained output format — and compare the three outputs against the original.
Check yourself
1. Why does adding one good example to a prompt often help more than three extra sentences of instructions?
2. What does "prompt versioning" mean, and why treat prompts like source code instead of throwaway text?
Tool use (function calling)
Answer the check-yourself questions to unlock this