Lesson 5: Positive vs negative instructions, and giving the model an 'out'
A pile of 'don'ts' is hard to follow: it tells the model what to avoid, but not what to actually do. A positive instruction — one that says directly what to write — is clearer and easier to follow. And instead of the model inventing an answer when information is missing, we'll learn to give it an 'o
Instead of listing what's forbidden ('don't be long, don't use jargon') — tell the model what to do ('write 2 sentences in plain language'). And when info might be missing, give it an out: 'if it's not in the text, write not specified'. That way it won't make things up.
- positive instruction
- An instruction that tells the model what to do, instead of listing a string of prohibitions ('don't').
- giving an out (escape hatch)
- Explicit permission for the model to say 'not specified' / 'I don't know' / ask a clarifying question when information is missing.
- hallucination
- An answer that sounds confident but is wrong or invented — the model fills a gap with a detail that wasn't in the text.
