Step-back prompting (Zheng et al., 2023) works because specific questions can trip models into surface-level pattern matching. Zooming out to the principle first forces the model to retrieve the right conceptual framework, then apply it.
Mechanically: "What is the underlying concept / principle / rule here?" → get the abstraction → "Now apply that to answer the original question." Two-pass, but one prompt. Especially effective on physics, chemistry, law, compliance, and policy questions.
Example Prompt
Original question: "Can an employer in California require employees to work during their meal break?"
Step back first:
1. What general labor-law principles govern meal breaks?
2. What exceptions typically exist, and on what grounds?
3. Who bears the burden of proof when a break is interrupted?
Now answer the original question, applying those principles.When to use it
- Specific questions in domains with governing principles (law, physics, compliance)
- The model has been surface-matching and getting it wrong
- You want the reasoning scaffold visible for audit / teaching
When NOT to use it
- Factual recall questions with no underlying principle to surface
- Creative tasks where abstract framing flattens the output
- Latency-sensitive tasks where the two-pass overhead matters