If you have grown accustomed to writing elaborate prompts for ChatGPT or Claude, complete with assigned roles, detailed instructions, examples and strict constraints, you may be doing more work than necessary. At least, that is the view of Boris Cherny, the developer behind Anthropic\u2019s Claude Code, who says improvements in AI models have changed how users should approach them.<\/p>\n
\u201cTalk to Claude the way you would talk to a coworker,\u201d Cherny wrote in a post on X. \u201cThere\u2019s no secret to prompting. There\u2019s no need to be overly scaffolded or prescriptive for most tasks.\u201d<\/p>\n
Rather than trying to find the perfect wording or anticipating every step a model should take, Cherny recommends focusing on three things: what you want the model to do, how much effort it should invest, and how it should verify that it has done the job correctly. The shift reflects his view that newer models can handle more of the decisions involved in completing a task.<\/p>\n
Cherny illustrated his approach by sharing a prompt he had used to ask Claude to create a visual artifact based on a podcast episode. He wanted the result to be visually impressive, comparing his expectations with the quality of data visualizations in The New Yorker or The New York Times. He also told the model to \u201cuse lots of tokens\u201d and floated the possibility of adding watercolor.<\/p>\n
That last suggestion was deliberately open-ended. When a user questioned whether Cherny wanted watercolor or not, he replied that he trusted the model\u2019s taste. The exchange offered a glimpse of the approach he advocates: communicate the desired result, provide direction where it matters, and allow the model some latitude over how to get there.<\/p>\n
The advice marks a departure from the prompt-engineering culture that took hold in the early days of generative AI. Users traded templates, specialized commands and elaborate formulas intended to coax better responses out of chatbots. Some prompts grew to hundreds or even thousands of words as people tried to control the model\u2019s behavior in advance.<\/p>\n
Cherny argues that this level of scaffolding is becoming less necessary for many tasks. He pointed to the difference between the Claude Sonnet 3.5 era, when precise wording mattered more, and newer models that can be given a goal and greater freedom to work out the details.<\/p>\n
But the simplicity of his example also prompted a debate about cost. One user joked about the instruction to \u201cuse lots of tokens\u201d in the context of unlimited usage. The comment points to a practical trade-off: allowing a model to spend more time and resources on a task may produce a better result, but it can also mean higher usage costs or more time spent waiting, depending on the service and plan.<\/p>\n
A short prompt is not necessarily a more efficient one, either. If a request leaves out important context or fails to define what a successful result looks like, the model may need several rounds of corrections. A more detailed set of instructions can save time and tokens when the task has specific requirements that the model would otherwise struggle to infer.<\/p>\n
Nor does Cherny\u2019s advice mean that users should stop giving models clear instructions. His three principles still require a defined objective and, importantly, a way to check the result. In software development, that could mean running tests. In other tasks, it might involve checking sources, comparing data or assessing the output against explicit criteria.<\/p>\n","protected":false},"excerpt":{"rendered":"If you have grown accustomed to writing elaborate prompts for ChatGPT or Claude, complete with assigned roles, detailed…\n","protected":false},"author":2,"featured_media":641298,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[85,46,125],"class_list":["post-641297","post","type-post","status-publish","format-standard","has-post-thumbnail","category-technology","tag-il","tag-israel","tag-technology"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/posts\/641297","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/comments?post=641297"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/posts\/641297\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/media\/641298"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/media?parent=641297"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/categories?post=641297"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/tags?post=641297"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}