Still Writing Prompts Casually? Top AI Companies Reveal Why It's Not That Simple

AI Study Club Report

Editor: Editorial Department

Introduction: YC reveals the important secrets of how top AI companies "tamed" LLM? 6 pages + prompt words, XML, and meta prompts are just the tip of the iceberg!

Recently, YCombinator invited representatives from many companies to explore how to improve the performance of artificial intelligence assistants through refined prompt writing. Disclosed how top AI startups “tamed” the big language model (LLM). In fact, their prompt word (Prompt) is always more than 6 pages, and they also use advanced operations such as XML tags and Meta-Prompting (see the repository below). Many people may think, "Isn't it just writing a prompt word? As for that complicated?" But I want to tell you: If you still have this idea, you may be lagging behind.

The prompt word seems simple, but it is actually a systematic project. Today, I will use my perspective and combine these "prompt word secrets" from YC to talk to you about how to do this well and truly make AI your right-hand man, not "artificial intelligence retarded".

1. The prompt word is not "written casually", it is the "communication language" between you and AI

When many people use AI, they write casually "Write an article" or "Assess the data", and then find that AI either answers the question or talks nonsense, and finally draws the conclusion: "This AI is not good, it's too stupid." But I ask you: Have you ever thought that it might be your own problem?

The prompt word is not an instruction that you can just type a few words. It is the communication language between you and AI. You have to give it clear, specific, actionable instructions just like you would for a new employee. How do top AI companies do it? Their prompt words are more than 6 pages long, how detailed is it? Task description, role settings, restrictions, output formats, and even possible mistakes must be listed. It's like writing a detailed "Onboarding Guide" to a new employee and telling him: Who are you, what you want to do, how to do it, and what level you can be considered qualified.

For example, you ask AI to write a marketing copy. You can't just say "write a copy". You have to tell it: You are my chief marketing officer, and the goal is to attract urban white-collar workers aged 30-40. The tone should be professional but not lose their affectionateness. The number of words should be controlled at 800 words. The format is "title + introduction + 3 key points + ending call". You cannot use too exaggerated words. If you don’t even explain these clearly, why do you expect AI to give you a perfect result?

So, before you use AI next time, ask yourself: Is my prompt word specific enough? If it is not enough, don’t blame AI for being “stupid” or for being lazy.

2. Give AI a "role" and let it know how to "play"

Whether people or AI, behavior patterns depend largely on "identity identity". If you give AI a clear character, it will be able to "get into state". YC's video mentioned that top AI companies pay special attention to "role propting". For example, they will tell AI: "You are an experienced customer support manager." In this way, AI will automatically adjust its tone and logic, prioritizing solving problems rather than throwing out a bunch of cold answers.

Let me give an example. I used AI to write a speech for me before. I told it directly: "You are my speech coach and have provided guidance to TED speakers. Now I want to prepare a speech on time management, for college students, with a humorous but illuminating tone, and control it within 10 minutes." What's the result? The manuscripts AI gave me are not only logical, but also have a little humor, which is particularly in line with my needs.

But if you don't set roles for AI, what will it do? It will answer in a "universal tone", which may be academic, or casual, and the final result is often thousands of miles away from what you want. So, giving AI a role is like giving an actor a script, so that it can "act" as if it is decent.

3. If complex tasks need to be disassembled, don't expect AI to be "in one step"

Many people like to throw a particularly big task to it when using AI, such as "written a 50-page book for me" or "watch me a complete business plan." Then the AI ​​either gets stuck or gives you a bunch of messy things, and you think AI is useless.

But I want to tell you: AI is not a god, it needs you to help it break the task into small pieces. YC's interview mentioned that top AI companies will break down complex tasks into predictable small steps to guide AI to "take every step". It's like when you teach a child to do math problems, you have to teach him addition and subtraction first, then multiplication and division, so that he can finally solve the equation.

For example, you want AI to help you write a book. You can first let it write an outline: list the chapter structure and the core content of each chapter. Then let it be written one by one, and each chapter is further divided into "introduction + subject + summary". By doing this step by step, the output quality of AI will be much higher.

In fact, simplifying complex problems is a must-have skill for all highly effective people. The same is true for AI. You have to help it break complex tasks into simple subtasks so that it can achieve its maximum value.

4. Use structured output, don't let AI's answers "in a mess"

Have you ever encountered this situation: the AI's answer is correct, but the format is messy, either the long and full of topics have no focus, or it cannot be used directly at all? YC's interview mentioned that top AI companies will use Markdown, XML tags or JSON formats to standardize AI output. For example, Parahelp will use <manager_verify> let AI generate answers in a fixed format.

I have tried this method and it works really well. For example, if I ask the AI ​​to help me organize a meeting record, I will tell it directly: "The output format is Markdown, and the structure is 'conference theme + participants + discussion points (using numbered list) + action plan (using table)'." As a result, the records given to me by AI are very clear, and I can use them directly.

If you don't give AI a clear output structure, it will answer according to its own "thoughts". In the end, you have to spend time sorting out, which is not worth the loss. Therefore, learning to use structured labels saves time and effort.

5. "Meta-tip": Let AI help you optimize prompt words

I especially like this method. YC's interview mentioned a technique called "Meta-Prompting", which means letting AI optimize your prompt words by itself. You feed your prompt words, the output of the AI, and the results you expect to another AI, and let it help you analyze the problems and make suggestions for improvement.

I tried it once, and the effect surprised me. I wrote a prompt word at that time, asking AI to help me write a product introduction, but the output of AI is always too "academic" and not attractive enough. I showed this prompt word and output to another AI, and asked it: "Where is the problem with my prompt word? How can I change it to make the output more attractive?" As a result, the AI ​​told me that I did not specify the tone and target user clearly. It suggested that I add the description "The tone should be lively and aimed at young people". After I changed it, the output of AI immediately became much better.

The core of this method is: AI can not only help you do things, but also help you "think". You have to learn to use its abilities and make it your "prompt word optimization master".

6. Evaluation is more important than the prompt word itself

Finally, I want to say something that many people ignore: Although prompt words are important, the more important thing is the "evaluation system" (Evals). Top AI companies will design a complete set of test cases for their prompt words, covering a variety of routine and edge scenarios, ensuring the reliability and stability of prompt words.

This reminds me of my experience in investing. Whether an investment strategy is good or not depends on how "smart" it is, but on whether it can withstand the test of the market. The same goes for prompt words. You have to test it in different scenarios to see if it really works. If not, you have to adjust until it can output the result you want.

I suggest that every time you finish writing a prompt word, you will prepare at least 5 test cases: 3 regular scenes and 2 edge scenes. For example, if you ask AI to write marketing copy, then test an ordinary product, a high-end product, or even a completely fictional product to see if the performance of AI is consistent. If it is inconsistent, it means that there is still something wrong with your prompt word.

Written at the end

Prompt Interning is simple on the surface, but it is actually a systematic job. It requires you to design carefully, continuously optimize, and even treat it like developing a core IP. The potential of AI is huge, but whether it can be realized depends on how you use it.