When you throw a specific, detail-heavy question at an LLM, it often latches onto the details and fumbles. Step-Back Prompting fixes this with one move: ask the general question first, then apply the answer to the specific. Here's a live side-by-side.
🪜 Watch direct vs step-back: https://dev48v.infy.uk/prompt/day17-step-back.html
The technique
Two steps:
- Step back — turn the specific question into a more general one. "If temperature triples and volume halves, what's the new pressure?" → "What principle relates pressure, temperature, and volume?" (the gas law).
- Apply — recall/retrieve that principle, then answer the original question using it.
The demo races a direct answer (fumbles the arithmetic) against the step-back path (states the law, applies it, gets it right).
Why it works
The general question surfaces the right framework before the model commits to a guess. It's the opposite failure mode from rushing: you give the model its footing first.
Where it shines
Science/physics, multi-hop reasoning, and especially RAG — a general query retrieves better background documents than a hyper-specific one. It pairs naturally with chain-of-thought (abstract up, then reason forward).
🔨 Full pattern (prompt for the step-back question → recall the principle → answer the original) on the page: https://dev48v.infy.uk/prompt/day17-step-back.html
Part of PromptFromZero. 🌐 https://dev48v.infy.uk
Top comments (0)