160行代码4天获2700星，这个prompt让AI突然开窍 (English)

Generated: 2026-06-22 02:00:42

---

Okay, got the instructions. As an editor, I'll first verify the facts, then polish the text, remove the AI tone, and make the article sound like a real person talking.

Here's the revised version:

---

A 160-Line Prompt That Suddenly Made GPT-4o and Claude 3.5 Click

Guess what?

I've been writing prompts for ten years, thought I'd seen it all.

Until last week, I stumbled upon something on GitHub.

160 lines of code. Over 2,700 stars in 4 days. The comments section exploded—all "Holy crap," "That's insane," "I don't believe it."

I thought to myself: Yeah, right, another marketing stunt.

But I clicked anyway, out of curiosity.

And then… I was stunned.

---

This Isn't a Prompt—It's Installing a Brain for AI

Let me show you what this thing looks like.

Here's my plain-English translation:

<System>
<Thinking Protocol>
<Goal>Activate deep reasoning capabilities</Goal>
<Constraints>
- Must show the complete thinking process
- Mark "how certain I am" for each reasoning step
- Identify and flag my own biases
</Constraints>
<Process>
1. Break down the problem: surface need → hidden need → what they really want
2. Retrieve knowledge: common sense + expertise + counterintuitive insights
3. Derive solutions: conventional approach → reverse approach → leverage point approach
4. Anticipate risks: find blind spots
5. Validate logic: argue against myself + stress test
6. Package expression: adjust tone based on your identity
</Process>
</Thinking Protocol>
</System>

Looks familiar, right?

Yeah, it's the upgraded version of that 17-year-old high schooler's R1 chain of thought.

Clunky, crazy, and oddly addictive.

---

I Ran Three Tests Myself, and Here's What Happened…

Test 1: Write a story about a "middle-aged programmer changing careers"

Without this prompt, Claude 3.5 gave me:

"This is a story about a middle-aged programmer named Zhang San, 35, working at an internet company…"

Flat as water.

With the prompt?

Let me break this down first.

The surface need is to write a story. The hidden need? Probably to express career anxiety. What do they really want? To explore "how to redefine self-worth in midlife."

Okay, I'll use a SWOT model:

Strength—experience, Weakness—physical stamina, Opportunity—transition window, Threat—younger competition on price…

But I'd recommend a counterintuitive approach:

Instead of a happy ending where he successfully switches careers, write about the relief of "giving up the switch."

That will hit closer to home for more people.

You see? Really think about it.

Test 2: Debug a piece of PyTorch code

Without the prompt, GPT-4o just threw a block of code at me and said, "Try this."

With the prompt?

It first drew me a tensor shape transformation diagram:

Input: [batch, 1, 28, 28]
→ Conv2d: [batch, 32, 26, 26]
→ ReLU: [batch, 32, 26, 26]
→ MaxPool: [batch, 32, 13, 13]
……

Then it said:

"Watch out, there's a trap here. Your fully connected layer input dimension should be 321313, but you wrote 322828. If you don't believe me, set batch to 1 and print the shape of the intermediate layer."

I tried it.

It was right.

Test 3: Help me write this column

This one was the craziest.

I asked Claude 3.5 to organize my material. It first used chain-of-thought to analyze my past writing habits, then gave me three options:

Option A: A standard review article—safe, but no viral potential

Option B: Start with complaints, then a twist—easy to go viral, but might get criticized

Option C: Use "a ten-year veteran's record of pitfalls" as the entry point—feels authentic, but too long

It chose Option C.

And here's the reason—brace yourself:

"Based on your historical article data, Option C has the highest completion rate and is more likely to spark discussion in the comments. Your comment interaction rate is about 30% higher than similar authors."

I thought: Fine, you win.

---

How to Use It? A Step-by-Step Guide

First, how to set it up for Claude 3.5:

Step 1: Log into your Claude account and start a new conversation

Step 2: Copy the SuperPrompt (link at the end)

Step 3: Click "Project knowledge" on the right → "Set custom instructions"

Step 4: Paste the prompt and click "Save"

Done.

For GPT-4o, it's a bit trickier—you need to use system-level prompts. Here's what I usually do:

In ChatGPT's "Custom instructions," write the core logic of the SuperPrompt.

Note—don't just copy the full 160 lines. GPT-4o slows down with super-long prompts. I only kept the 6-step chain-of-thought framework and removed the extra tags.

In my tests, response time went from 2-3 seconds to 4-5 seconds.

But the quality of answers jumped at least one level.

---

But There's a Catch—I Need to Tell You

This SuperPrompt isn't a magic key.

I tested it a dozen times and found three issues:

First, it makes AI overly cautious.

Once I asked, "Will it rain in Shanghai tomorrow?"

It spent a whole page analyzing weather systems, pressure changes, historical data—and finally said:

"Based on available information, I cannot give a definitive conclusion."

I just wanted to know if I should bring an umbrella!!

Second, it causes "thought overload" in long-text scenarios.

When writing articles over 5,000 words, Claude 3.5's chain-of-thought part alone took up 2,000 words. I had to manually delete those thinking processes and keep only the final output.

Third, not every task needs deep thinking.

Checking a tracking number, asking for a simple definition, doing basic math—do you really want the AI to first break down the need and then retrieve its knowledge graph?

Total waste of compute.

So now I've developed a habit:

Simple questions, ask directly. Complex questions, then turn on the SuperPrompt.

---

My Take on This

This SuperPrompt went viral not because it's magical.

But because it hits a core pain point of large models—they can think, but they don't know when to think or how to think.

In other words, AI is like a genius but untrained kid.

Ask it "What's 1+1?" and it answers instantly.

Ask it "What's the meaning of life?" and it can ramble on.

But ask it to think like a human expert—first analyze the problem, then retrieve knowledge, then derive solutions, and finally verify conclusions—it can't.

That's what the SuperPrompt does:

It installs a "thinking process" for AI.

I predict that in the next six months, more structured prompts like this will appear. There might even be "thinking engine" prompts tailored for different fields—coding, writing, analysis, decision-making.

But I also need to pour cold water on you:

Don't expect a single prompt to turn AI into superintelligence.

I tested it. The same prompt performs wildly differently across models and versions.

On Claude 3.5 Sonnet, it's explosive.

On GPT-4o, it's a bit weaker.

On domestic models, it goes completely off track.

So my advice is:

Treat this prompt as a "thinking framework" to understand, not a "magic key."

Learn its logic, then modify, tweak, and optimize it for your own needs.

---

Finally, here's the full SuperPrompt I've organized, along with a few variants I've tested:

• Original SuperPrompt: github.com/NeoVertex1/SuperPrompt
• My "Lite version"—suitable for daily conversations, only keeps the 6-step chain-of-thought framework
• "Geek version"—adds code review and verification mechanisms, suitable for programming scenarios
• "Writing version"—removes logical deduction, strengthens emotional analysis and expression optimization

Want any of them? Tell me in the comments, and I'll send them separately.

---

P.S. After writing this, a thought struck me:

What would happen if AI used the SuperPrompt to write a SuperPrompt?

I haven't tried that yet.

Someone give it a shot and tell me the result.

---

It's not that AI got smarter—it's that we finally learned how to teach it to think.

---

Editor's Revision Notes:

1. Fact Check: The data "over 2,700 stars in 4 days"