What is step-back prompting?

A technique from Google DeepMind research where you first ask the model a more abstract, higher-level question to surface relevant principles, then ask the specific question. Grounding in the general case measurably improves accuracy on reasoning-heavy tasks.

When does it help most?

On questions that hinge on a principle, formula, or general fact — physics, chemistry, law, multi-step reasoning. It helps less on simple lookups or creative tasks where there is no underlying rule to recall.

Why answer the general question first?

Models reason better when the relevant high-level facts are already in context. Deriving the principle before the specifics prevents the model from anchoring on surface details and skipping the rule that actually governs the answer.

Does this tool run the prompt?

No. It builds the two-stage prompt for you to paste into any LLM. The builder runs locally in your browser and only outputs text.

How do I pick the abstraction level?

Step back just far enough to reach the governing rule. Too little and you restate the question; too far and the principle becomes irrelevant. The builder offers principles, category, and definition levels to match common cases.

Step-Back Prompting Builder

Step-back prompting builder

Step-back prompting is a reasoning technique: before answering a hard specific question, you ask the model a more general one to surface the underlying principle, then feed that grounding back into the specific answer. Research from Google DeepMind showed it improves accuracy on knowledge-heavy reasoning tasks. This builder turns any specific question into the correct two-stage prompt.

How it works

You provide the specific question and its domain, and pick how far to step back. The builder generates an abstraction question matched to that level — toward fundamental principles, a broader category, or a general definition — then assembles a prompt that instructs the model to (1) answer the general question first, stating the relevant principles, and (2) use that to answer your specific question. The two stages share one context so the model reasons from fundamentals rather than surface cues.

Tips and example

Step back to the rule, not the topic. For “Why does ice float?” the step-back is “What determines whether a solid floats in its own liquid?” — the density principle — not “Tell me about water.”
Keep the domain accurate. It sharpens the abstraction question.
Best for principle-driven questions. Physics, chemistry, finance, and law benefit most; trivia lookups benefit least.
Chain, don’t split. Keeping both stages in one prompt lets the model carry the derived principle straight into the answer.