In a 2023 test, GPT-3.5 did worse with the answer buried in the middle of twenty documents than with no documents at all.
L5 · Talking to Models
A prompt is the opening of a document, and the model continues it. Two things shape what comes out: the text you hand over, and the dial that turns scores into chances. You will see why two worked examples beat a paragraph of adjectives, when step-by-step working helps, and how to get output that always parses. Then the hard parts: made-up facts, prompt injection, testing a prompt like code, and knowing when to stop prompting and change something else.
- Where a Prompt Goes
- Three Voices, One Document
- Say the Thing You Want
- Show, Do Not Tell
- Room to Think
- One Dial Called Temperature
- Cutting the Tail
- Output You Can Parse
- Lost in the Middle
- Why Models Make Things Up
- Prompt Injection Is a Security Bug
- Test Sets, Not Vibes
- When Prompting Is the Wrong Tool