And a Harvard study confirmed it. The 3 eras of prompting — and why reasoning models want something completely different from you.
8 min read
Just now
--
Press enter or click to view image in full size
On March 25, 2025, Google published a documentation page for the Gemini 2.5 thinking model API. It landed on the ai.google.dev developer portal between configuration reference tables and sample code. Most developers never read it past the thinking_budget parameter definition.
Buried inside was guidance that quietly invalidates two years of mainstream prompting practice.
The key line: for models that already reason internally, adding chain-of-thought instructions to your prompt — “think step by step,” “reason through this carefully,” “first analyze, then conclude” — does not help. In many cases, it actively interferes. The model is already thinking. You are not improving its thinking by narrating it back to the model. You are adding noise to a process that has already started without you.
A month earlier, Anthropic’s updated Claude documentation had said something structurally identical. Their guidance for Claude Opus 4.7 now warns: if you find the model thinking more often than you’d like, it can…