Does adding "think step by step" to prompts actually help?
Now that reasoning models are coming out, I've heard people say that phrase is no longer necessary—I'm curious what your experience has been.
I still break things down for complex tasks, like "first make a plan, then execute it." That seems to clearly give better results.
If anyone has seen data showing it's actually counterproductive with the latest models, please share.
6 answers
Agreed — when I give it something complex, I always break it into steps lol. If you just throw it all at once, it often goes off track halfway through.
Reasoning models and regular models need to be treated somewhat differently. For o1 or R1-series models, since they already go through an internal reasoning process, adding phrases like "think step by step" often just duplicates the instruction and wastes tokens—this has also been mentioned in official guides. Conversely, when giving complex tasks to GPT-4o or small models, it still feels better in practice to break things down into planning and execution. In the end, it seems right to view it as depending on the model type and task difficulty.