Step-Back Prompting
Tier: Advanced · Task types: Abstract reasoning, concept explanation, first-principles analysis · Difficulty: ⭐⭐⭐☆☆
Table of contents
Definition
Step-back prompting asks the model to first answer a more abstract, general version of the question, then use that general answer to reason toward the specific answer. It forces the model to activate broader conceptual knowledge before narrowing to specifics — improving accuracy on tasks that require domain knowledge or first-principles reasoning.
The technique involves two prompts: (1) “What are the general principles governing X?” and (2) “Given those principles, answer this specific question: Y?”
Use Cases
- Physics / science problems where the correct answer requires applying a general law to a specific scenario
- Policy analysis — stepping back to general principles before evaluating a specific case
- Debugging — identifying what category of problem applies before diagnosing a specific bug
Weak Prompt (Direct Question)
A 2kg ball is dropped from a height of 20 metres. What is its velocity
just before it hits the ground? Ignore air resistance.
Output
Using kinematics: v² = u² + 2as = 0 + 2(9.8)(20) = 392. v = √392 ≈ 19.8 m/s
(Correct, but the model may struggle on less standard variants without the conceptual grounding.)
Improved Approach (Step-Back)
Prompt 1 — Step Back:
What are the key physics principles that govern how objects fall under gravity
and how to calculate their velocity at impact? Explain the core equations.
Output 1:
Energy conservation: PE = KE at impact → mgh = ½mv² → v = √(2gh). Kinematic alternative: v² = u² + 2as, where u=0, a=9.8 m/s², s=height. Both approaches give the same result for free-fall problems.
Prompt 2 — Apply to Specific:
Using those principles, calculate the velocity of a 2kg ball dropped
from 20 metres just before impact. Show which equation you're using and why.
Output 2:
Using energy conservation (v = √(2gh)): v = √(2 × 9.8 × 20) = √392 ≈ 19.8 m/s. Note: Mass is irrelevant — the 2kg is a distractor; free-fall velocity depends only on height and g.
Assessment: The step-back correctly identified that mass is a distractor — something the direct prompt may not explicitly surface.
Verdict
Step-back prompting is particularly effective for problems that contain irrelevant information (distractors) or where the model tends to pattern-match to surface features rather than underlying principles. The two-prompt overhead is worth it for high-stakes reasoning tasks.
Common Pitfalls
1. Using step-back on simple factual lookups “What is the capital of France?” does not benefit from first-principles reasoning. Reserve for genuinely complex tasks.
2. Not connecting the step-back to the specific question The second prompt must explicitly reference the principles generated in the first. Don’t assume the model will connect them automatically.
3. Accepting a shallow step-back If the first prompt produces a vague or surface-level abstraction, the second prompt will inherit that shallowness. Push for depth: “What are the fundamental equations?” not “What is important about this topic?”