Claude Code creator says most users are making one mistake with AI prompts

Claude Code creator Boris Cherny has revealed that most users may not be getting the most out of Anthropic’s popular AI chatbot because of one common mistake. Speaking at a recent Y Combinator event, Cherny explained that many users micromanage AI models with lengthy prompts, which he called a “common mistake” because it prevents the AI from determining the best way to accomplish a task.

“I think a really common mistake that I see is people are using Claude Code, they’re using Claude, and they just give it way over-specific instructions,” Cherny said. “They’re like, ‘I want you to do this, but I want you to do it in this way, this way, this way. You must do like one, then two, then three, then four.'”

Cherny argued that while this approach may not have worked well even six months ago, things have changed significantly with the latest generation of AI models.

He also argued that users should give AI models “harder tasks” than they think they can accomplish. Cherny said this works because every new model generation is capable of undertaking tasks that the previous one could not.

The Anthropic executive gave the example of how Claude recently rewrote an entire production codebase from one programming language to another over an 11-day period. According to him, the same task would have taken engineers more than a year to complete.

Prompt engineering is not the future:

While there has been a lot of discussion around learning prompt engineering to get the best out of AI models, Cherny argued that the future is actually about giving AI models more autonomy to accomplish tasks on their own.

“The verification I think is probably the single most important thing that people do not get right,” he said, explaining that users should provide AI with ways to test and validate its output instead of relying on increasingly elaborate prompts.

“I think people tend to overthink it a little bit. I think people tend to overengineer,” Cherny said. Rather than trying to force AI to follow the exact process a human would use, he suggested treating it “like you would a coworker,” saying that is “the level of intelligence that it’s at now.”

Cherny is not the only person to argue against giving AI models overly specific prompts. In fact, Google Brain co-founder Andrew Ng shared a similar idea in a post on X (formerly Twitter) last year, which he called “lazy prompting”.

Under this approach, Ng said users should begin with a short, imprecise prompt and only add more context if the model’s initial response isn’t good enough. He explained that this approach works well when users can quickly judge whether the output is correct, such as while debugging code or editing text.

“I find it’s sometimes faster to be lazy and dash off a quick, imprecise prompt and see what happens,” Ng wrote. He added that instead of writing lengthy instructions upfront, users should “add details to the prompt only when they are needed.”