Claude Code: Boris Cherny Recommends Simple Prompts

Boris Cherny, the creator of Claude Code at Anthropic, advocates for direct instructions rather than complex prompt engineering. He illustrates this approach with an example that sparked a discussion about expected accuracy and expresses trust in the judgment of Opus 5.5.

A Debated Example and a Model Considered Reliable by Its Creator

In one of his prompts, Boris Cherny used the phrase "Maybe add watercolor," which led a user on X to suggest he be more assertive in his request. Cherny responded that he trusts the discernment of Anthropic's Opus 5.5, stating that it "has good taste" and that he relies on it to "cook." He also shared the result obtained: Claude generated several watercolor-like graphics in beige and orange hues.

Goal-Oriented Instructions, Effort, and Verification

Boris Cherny advises formulating prompts around three axes: specifying what is expected from the model, indicating the desired effort, and requesting verification of the result. He recommends addressing Claude as a colleague, without overly structuring or prescribing for most tasks. According to him, there is no special secret to interacting effectively with AI.

An Evolution from Yesterday's Prompt Recipes

Boris Cherny believes that the improvement of models has reduced the need to refine prompts, unlike a recent period when users exchanged template formulations to guide chatbots. Claude from Anthropic was often associated with the instruction "Don't make mistakes." Today, Cherny formalizes his advice into three points and shares an example of a prompt in which he describes a podcast episode as "really amazing," requests the creation of an artifact, the use of numerous tokens, and visuals surpassing those of the New Yorker or the New York Times. He is the creator of Claude Code for Anthropic.