Using AI, just think like it has dementia

Writing code for AI is an important skill. Photo: Toptal .

One of the most essential office skills today is writing commands for chatbots to achieve the desired results. This may sound simple at first, but it is quite time-consuming and laborious if you don't know how.

AI company Anthropic recently published a white paper called “Prompt Engineering Overview” to help beginners get started. While it can be applied to most chatbots, it was designed to optimize the experience with Claude, Anthropic’s chatbot.

According to the document, the first thing to do is to understand the nature of the chatbot. “When interacting with Claude, imagine that he is a very smart new employee who has temporary amnesia and needs specific guidance,” the company explains.

Next, users should have a rough idea of what their query is asking, along with a visual of what they want the results to look like. Anthropic also provides a tool to help you create an initial draft. The most important thing is to refine the query.

The document above provides some tips for effective commands. First, the chatbot is like a child, only carrying out the request. Since it does not know the user’s preferred working style or norms in advance, the more specific the explanation, the better Claude will respond.

Information that should be made clear to the chatbot up front includes what the results will be used for and for whom, as well as the ultimate goal of the task. Anthropic also recommends presenting requests as bulleted or numbered lists to help organize them.

To save time and get the results you want, provide your chatbot with examples. “By including a few well-crafted examples, you can significantly improve the accuracy, consistency, and quality of Claude’s responses,” says Anthropic.

This strategy is sometimes called multi-shot prompting. Anthropic says it helps reduce misunderstandings and ensures the results have a consistent structure and style.

For complex requests, giving the chatbot ample time to think can dramatically improve performance. Anthropic calls this the chain-of-thought (CoT) technique, which encourages Claude to analyze the problem step by step, leading to more accurate and in-depth answers.

Using this approach allows users to get the most out of the sequence of steps they need to take, and also allows the chatbot to think sequentially before responding. “This thoughtful reasoning process leads to more confident and well-founded responses,” Anthropic explains.

Another effective trick is to assign a chatbot a specific role, such as a news editor or a financial expert. This works best in complex situations like legal analysis or financial modeling, ensuring you get the style and content you want.

Finally, be aware that chatbots sometimes make up information, so you should always double-check any results they give you. Anthropic says the most effective way to reduce “illusion” is to allow chatbots to admit when they don’t know the answer.

“Make it clear to Claude that it is okay to admit uncertainty. This simple technique can dramatically reduce the likelihood of providing false information,” Anthropic explains. Users can also ask Claude or other chatbots to cite sources for the information they provide.

“You can ask Claude to validate each argument by finding a supporting citation after it generates an answer. If it can't find a citation, Claude is forced to retract that information,” Anthropic instructs.

Source: https://znews.vn/dung-ai-cu-nghi-nhu-no-bi-benh-mat-tri-post1570435.html