When zero-shot wins
Standard tasks: summarization, translation, generic classification, basic Q&A. Models have seen these thousands of times in pretraining.
Advertisement
When it fails
Custom classification schemas. Domain jargon. Output formats not seen in training. Ambiguous instructions.
Advertisement
Signals it's not working
Unstable output structure across runs. Model asks clarifying questions. Response includes 'I'm not sure' hedging.
Upgrade path
Zero-shot → few-shot (add 2-5 examples) → chain-of-thought (add 'think step by step') → agentic (add tools).