For a long time, better results from language models seemed to require longer prompts. More context, more explanation, and more detailed specifications helped compensate for limited autonomy and narrow task representations, particularly when systems were brittle and benefited from external scaffolding. As models have improved, that relationship has inverted, with long prompts increasingly degrading performance rather than improving it. They inject assumptions, lock in flawed premises, and force the model down human-shaped reasoning paths that are narrower than the problem itself. As a result, giving the model less context often leads to better outcomes.
Why This Changed
Earlier models relied on the human to do much of the thinking, with step-by-step instructions reducing error because the model could not reliably plan on its own.
Modern models behave differently. With larger context windows and stronger internal planning, they can build the relevant context as they work, which means that front-loading a prompt with explanations fills that working memory with human framing before the model has a chance to reason independently.
The Cost of Explaining Too Much
Detailed prompts encode beliefs as well as instructions, since potentially incorrect assumptions become constraints the model might not question. The system optimizes for the description rather than the underlying goal, often appearing competent while quietly solving the wrong problem.
Providing less initial input helps avoid that trap by allowing the model to decide what information matters, surface uncertainties, and construct its own representation of the task.
The Human Contribution
As models get better, the human contribution shifts away from explanation and toward judgment, with the highest leverage coming from stating the problem clearly, answering clarifying questions, and evaluating outputs against reality.
Over-prompting reflects habits formed during an earlier phase of model capability, and as that phase ends, restraint increasingly becomes a competitive advantage.
Subscribe for new essays
Get updates when new writing goes live.
