Stop Overthinking Prompts: Claude Code Developer Advises Natural AI Dialogue
Boris Cherny, developer of Claude Code, argues that advanced AI models make complex prompt engineering obsolete, advising users to talk to models like colleagues and focus on clear goals and verification.
If you have grown accustomed to crafting long prompts for ChatGPT or Claude featuring defined roles, rule lists, steps, limitations, examples, and detailed guidelines, you might—according to someone who knows these models intimately—be working much too hard. Boris Cherny, the developer who built Claude Code, published a post on X this week proposing a much simpler approach to working with AI: "Talk to Claude as you would to a colleague." According to Cherny, for most tasks there is no longer any need to build a detailed framework for the model or dictate precisely how it should execute every step. This shift, he explains, stems from the continuous improvement of the models.
The Shift in Prompt Engineering
Back in the era of Claude 3.5 Sonnet, Cherny wrote, exact prompt wording carried far greater importance. Today, he argues, instead of spending time trying to find "magic words," it is crucial to convey three main elements to the model: what exact action you want performed, how much effort you expect it to invest in the task, and how it is supposed to verify that the final result is correct. Cherny's post was prompted by surprised reactions after he shared a prompt he had used himself. He asked Claude to build an interactive website based on provided content to visually and engagingly present a story and its core insights. Instead of breaking the task down into dozens of detailed instructions, he relied on relatively general guidelines, specified the effort level, and left the model the freedom to make some decisions independently.
"Talk to Claude as you would to a colleague," suggested Boris Cherny, noting that advanced AI models no longer require hyper-detailed prompt engineering for everyday tasks.
Efficiency and the Token Debate
This approach largely contrasts with the "prompt engineering" culture that emerged at the dawn of generative AI. Users used to collect templates, commands, and techniques meant to squeeze better answers from models, sometimes building prompts hundreds or thousands of words long. However, many commenters poked fun at Cherny for including the instruction "use a lot of tokens" in his prompt. One user joked: "It's funny to write 'use a lot of tokens' when you have unlimited usage." Behind the humor lies a valid critique: it is easier to rely on a brief, general prompt when you can simply let the model keep working, retry, and burn tokens until it hits the right outcome. For regular users, however, a slightly more detailed prompt can actually save unnecessary iterations and conserve valuable tokens.
It is important to emphasize that Cherny's advice does not mean "writing less" in every situation. The three rules he proposes still require the user to clearly define the goal—and particularly to give the model a way to verify its work. In programming tasks, this might mean running automated tests; in other tasks, it could involve source checking, data comparison, or evaluating the outcome against predefined criteria.