Chain of thought is dead. The principle isn't.

OpenAI now tells you not to use the prompting technique that put “chain of thought” into every AI playbook on LinkedIn.

Their reasoning best practices docs: “Avoid chain-of-thought prompts: since these models perform reasoning internally, prompting them to ‘think step by step’ or ‘explain your reasoning’ is unnecessary.” Google deprecates it in favour of native thinking mode. Anthropic calls it a fallback for when thinking is switched off.

Three providers, same message. The technique is a vendor anti-pattern on reasoning models.

The receipt

Chain-of-thought prompting was peer-reviewed gospel. The landmark paper landed in 2022, and within months “think step by step” was the most-repeated advice in AI. It worked, too. On the models available then, it produced measurable gains on reasoning tasks.

The Wharton Prompting Science Report 2 (Meincke et al., June 2025) shows where that stands now. On reasoning models, CoT produced gains of 2.9%, 3.1%, and -3.3%. That last number is negative: it made the model worse. Every condition cost 20 to 80 percent more time.

CoT still works where the model doesn’t already think for itself. The same study found a 13.5% improvement on Gemini Flash 2.0 and 11.7% on Sonnet 3.5, both non-reasoning models. What died is the universal advice. “Always add ‘think step by step’” is now wrong in exactly the places where it used to matter most.

The implementation collapsed. The reason it worked didn’t.

What the principle actually is

CoT worked because it forced decomposition: break a big problem into smaller steps, review each output before the model continues. That’s an engineering pattern, not a prompting trick. Decompose the work, gate each stage with a review. It shows up everywhere serious AI systems get built, from multi-agent pipelines to production workflows.

Most founders never looked at why it worked. They grabbed “think step by step,” pasted it into their prompts, and moved on. They had the tactic without the principle. When reasoning models stopped needing the tactic, they had nothing underneath.

The principle: break work into stages, review each output before feeding it to the next. That structure works whether the mechanism is chain-of-thought prompting, native model reasoning, or a multi-agent pipeline.

What this means for you

If you adopted “think step by step” because someone on LinkedIn said so, you’re about to adopt the next thing someone on LinkedIn says. And the next. Each time, you rebuild from scratch.

If you understand that the pattern underneath was decomposition and review gates, you don’t rebuild. You swap the mechanism. The reasoning model handles the thinking natively now, so you stop prompting for it and spend that effort on better stage boundaries and review criteria instead. Ten minutes of adjustment, not a quarter of rework.

This is the question worth asking about any AI technique: if it died tomorrow, would your setup still work? If the answer is no, you’ve built on a tactic. Tactics have a shelf life of about six months in this space.

I wrote about this pattern in Why AI Setups Rot. The setup, which model, which prompts, which file structure, is always disposable. The layer underneath, your workflows, your voice, your operating context, is what compounds. Build on the layer that lasts.

If you want that layer installed properly in your business, that’s what the AI Brain engagement does. Or if you’d rather talk it through first, book a call.