Broad vs Narrow Prompting

Prefer task-agnostic prompts that elicit broad reasoning skills when you need one template across tasks; reserve narrow per-task templates for format-heavy niches.

Most prompts are narrow: few-shot by construction, or zero-shot templates engineered per task. They elicit task-specific skills. Broad prompts are multi-task cues that aim at a cognitive ability rather than a dataset — zero-shot chain of thought’s fixed “Let’s think step by step” is the canonical example for system-2 / multi-hop reasoning.

Harness guidance: start with a broad reasoning trigger (and measure under reasoning trigger sensitivity) before authoring per-task exemplar packs. Broad prompts are orthogonal to instruction tuning — they still help InstructGPT, vanilla GPT-3, and PaLM-class models — so do not assume chat-tuned models make CoT triggers redundant. Keep narrow prompts when answer format or domain demos are the bottleneck (cross-task few-shot CoT transfer; when chain of thought helps). Version both kinds in the prompt catalog.

---

Sources

• Large Language Models are Zero-Shot Reasoners (Kojima et al.), §5