Prompt Word Engineer Confession: How I use a trick to liberate my productivity

“In the symphony of the AI ​​era, the prompt engineer is the silent but crucial conductor, using carefully woven language to guide the intelligent behemoth of the big model to create the miracle of human thinking.

Since the birth of the new production paradigm, I have added a new identity to my job - prompt engineer.

I shuttled between different large language models, seeking resonance, and studied various prompt word paradigms, trying from few-shot, COT, self-consistency to dynamic prompt words. I once installed an x ​​to do a prompt word engineering training for the company, trying to transfer the pain and happiness of writing prompt words to everyone.

However, after dealing with it for a while, I found that there is still a language barrier between people and the big model - you don’t know me and I don’t know you.

In fact, we usually use more domestic models, especially before DeepSeek was born. Zhipu and Qianwen were used in production environments. The various prompts for chatgpt claude circulating on the Internet are really not that applicable. I often spent half a day and still couldn't generate the desired effect.

My friend next door once said with emotion: "After spending half a day chatting with LLM, I finally felt that we were on the same wavelength."

Sometimes I have to write more than a dozen tasks a day (yes, that's the intensity, I'm not afraid of rushing out a system in two or three days). I get several chicken legs a day.

Finally one day, I couldn't stand it anymore and I rebelled. I wanted to get rid of these days of being enslaved by big models!

Let AI generate prompt words

Previously, a team in Singapore proposed the CO-STAR framework, which has a good idea, but there are still too many words to write. As for the prompt word issue, the large language model can solve it by itself!

The following is a useful prompt word structure that I frequently use in multiple projects :

1. Role

The big model needs to play an identity or position.

2. Objective

What is the core objective of the mission?

3. Constraints

What are the hard requirements or restrictions?

4. Workflow

What steps should be followed to execute the task?

5. Example

Provide specific examples to ensure that the output style meets expectations.

Based on the above structure, I created a tool to automatically generate coze prompt words. I provide it to everyone for use (the qwen-max model is used by default)

https://www.coze.cn/store/project/7484301292970672179?from=store_search_suggestion&bid=6fnco0ods7g1j 

Just enter the task you want to perform in the task description. For example, "Enter a user question and generate the corresponding MYSQL query statement." Then click [Button] (what a straightforward logo)

The process behind this is as follows

Will it be expensive to frequently call on large models?

It seems that the large language model is frequently called in the process, but don’t forget that the billing of the large model is calculated based on tokens (simply understood as the number of words). In fact, although there are many processes, there are not many tokens.

Additionally, having large models focus on one task at a time often produces better results.

It turns out that being loyal is indeed more charming.

Let’s see it in action

The following case explains why the prompt word generation tool is more useful.

Enter a task description:

Last words

At the boundary of AI and human communication, prompt engineering is not only a translator of language, but also a transmitter of thought. When we learn to let AI write for AI, we actually enter the field of meta-prompt engineering, truly playing with technology, rather than being played by technology. Prompt word engineering may eventually die out, but at present, the ability to think about how to guide machines to think is still a core quality that everyone must have to improve productivity in the AI ​​era.