Zero-Shot Prompting
Zero-shot prompting is a prompting style that asks a model to perform a task with no examples, relying on instructions and the model’s prior training.
In plain English
Zero-shot means no examples. You describe the task in words, and the model does it. "Classify this review as positive or negative" with no training examples attached is zero-shot prompting. Few-shot would be the same instruction with three labeled examples included. Fine-tuning would be training the model on thousands of examples.
Frontier models in 2026 are strong enough that zero-shot works for most tasks, which is why modern agent systems default to it. Examples only appear when a task is unusual enough that the model needs to see the format — even then, one or two examples usually suffice. The cost savings are real: fewer tokens per request, less prompt engineering overhead.
Why it matters for Black Box
Black Box specialists are mostly zero-shot. Their system prompts describe the role, the tools, and the output shape; they rarely embed labeled examples. When a task is unusual — a niche Skill Pack, a non-standard format — the Skill Pack adds a few-shot example to the playbook.
Examples
- "Summarize this email in one sentence." — no example needed.
- "Classify this lead as cold, warm, or hot based on this context." — zero-shot unless the taxonomy is fuzzy.
- "Generate alt text for this image." — works zero-shot on modern multimodal models.