Original video in Chinese.
Key Takeaway
- The essence of prompting is general communication skill, not an AI-specific technique; its foundation is logic and methodology.
- Learning prompting should focus on improving logical thinking, communication methodology, and writing ability, rather than buying prompt courses.
- Good prompts need to be optimized for different AI models’ “cognitive patterns,” which requires repeated trial and comparison to understand.
- For ordinary users, mastering the underlying logic of prompts is enough, without going deep into model optimization.
- The article criticizes the “low ceiling” of prompt engineering and the phenomenon of being idolized by beginners, and encourages pursuing a higher level of cognition.
A lot of people have recently asked me how to learn prompt engineering. In particular, whether they should buy those prompt courses. I thought about it, and I’ve made almost a hundred AI videos, but I really haven’t talked about prompting. So this episode counts as a general reply.
Let me start with the conclusion: don’t learn prompting, and definitely don’t buy those courses. They’re useless.
Prompting has two layers. At the very bottom are logic and methodology. As I’ve said in Knowledge Planet, this layer has basically nothing to do with AI.
Think about it: what is the essence of prompting?
In one sentence: using a clear logical structure and expression to convey tasks or questions to AI.
The user provides a clear goal, context, and constraints, and ultimately guides the model to produce the expected result.
So, to put it bluntly, prompting is just organizing language to drive model reasoning. It is actually a general communication skill, very similar to task instructions between humans.
Let me give you an example, and you’ll understand.
I asked Claude an opinion-based question: What’s your take on Real Madrid’s performance this season?
After Claude searched for information using MCP, it answered using the “total-part-total” structure we’re familiar with:
At the beginning and the end, it mentioned that Real Madrid is still a top club and remains highly competitive; in the middle, it listed the achievements, and of course also some setbacks.
Setting the conclusion aside, this answer doesn’t seem to have any major problems. Most people, when expressing an opinion, would already be doing very well if they could manage “total-part-total.”
But if your logic is good and you understand some methodology for expression, then the AI’s answer will be very different.
I opened a new window. This time, I gave Claude a set of methodology I use most often.
Use the Point-Reason-Example-Point logic to answer.
- Point: state the point first;
- Reason: give the reasons;
- Example: based on the point and reasons, give corresponding examples;
- Point: restate the point at the end.
Same question: What’s your take on Real Madrid’s performance this season?
You can see that this time, Claude’s answer is much clearer.
First, it gives a clear point: Real Madrid has underperformed across multiple fronts.
Then it gives the reasons: injury problems, an incomplete tactical system, losses in key matches against Barcelona, and the coexistence issues among several core players.
Then it gives a bunch of examples: for instance, they failed in the Champions League, La Liga, the Copa del Rey, and the Super Cup.
Finally, it restates the point: Real Madrid ended this season empty-handed, and there are problems everywhere.
The prompt I just gave—Point, Reason, Example, Point, or PREP—is a methodology specifically used to organize and express opinions.
So, does this PREP count as “prompt engineering”? Yes and no.
Yes, because I really did put it into the prompt and made use of prompting.
No, because it isn’t some unique thing made specifically for AI. I use it when writing scripts. I also unconsciously use it when I’m speaking normally. I simply taught it to AI so its opinionated expressions would be more logical. That’s all.
There are many similar examples. Once you break apart those flashy prompts and study them carefully, you’ll find that briefing a model is actually the same as briefing your colleagues or suppliers. Like I said earlier, it really is just a general communication skill.
If, like me, you have strong logic and know how to distill methodology, then there’s absolutely no need to specifically learn prompt engineering. I skip anything related to this directly.
If your own logic is not that strong, then does learning prompts help? Yes, but not much. Because without logical ability as support, it’s hard to transfer what you learn to other situations, and you can only use prompts others have made in limited scenarios.
That’s the awkward part of prompt courses today. People who understand don’t need them; people who don’t understand can only use them to that extent anyway. And course sellers don’t dare teach the most basic, most essential things. Because once that’s exposed, everyone will realize it doesn’t have much to do with AI. Once the AI gimmick is gone, how are you supposed to make money?
That’s why I suggest the friends in the community not buy any prompt courses. If you want to learn, go learn logic, learn communication methodology, learn writing methods, and so on. All of that is fine. But don’t deliberately go learn prompting.
OK, that’s the bottom layer of prompting; everyone understands it now. Then the next layer up is optimizing for models.
Different AI models have different sensitivities to prompts, different ways of understanding them, and different response styles. For example, some models need very specific instructions, while others are better at handling open-ended questions.
So good prompts need to be optimized for the model’s “cognitive pattern.”
Then how do we know the “cognitive patterns” of various models? Unfortunately, the only way is through repeated trial and comparison, gradually understanding the model’s behavior patterns.
For example, context sensitivity: see whether the model omits key details; instruction-following: see whether the model ignores subtle instructions; creativity: see how much imagination the model has.
For ordinary users, there is no need at all to get into this layer. You only need to do the most basic, bottom-layer part well. Like the example I just showed, the model’s opinion expression will be much better.
If you really plan to harvest traffic through prompts, then spend some time getting into the second layer. Becoming a so-called “prompting guru” is not hard at all. It’s just pretty boring. First, because the ceiling is too low; second, because you’re worshipped all day by a bunch of beginners.
Being a big fish in a small pond is nothing to be proud of. Getting out of the pond is the real priority.
OK, that’s all for this episode. If you want to understand AI, want to become a super-individual, and want to find like-minded people, come join our newtype community. See you next time!