我花了两天实测李继刚的汉语新解Prompt，证实Claude 3.5确实断层领先 (English)

Generated: 2026-06-21 02:41:51

---

After Testing This God-Level Prompt, I Really Got Slapped in the Face by Claude 3.5

Let me tell you something straight.

Six months ago, someone told me "Claude 3.5 is head and shoulders above everyone else." I laughed out loud on the spot.

How could that be? In this AI battlefield where everyone's chasing each other, today you're number one, tomorrow someone else is. Nobody stays on the throne for more than three months. GPT-4o updates every other day, Gemini's scrambling to catch up—where's this "head and shoulders above" coming from?

Fast forward to a few days ago. Li Jigang's "Chinese New Interpretation" prompt exploded all over my WeChat Moments again.

I thought, fine, bring it on. Let's see if you're a horse or a mule.

You know what happened next?

My face swelled up so bad even my own mother wouldn't recognize me.

---

Before We Talk About This Prompt, Let Me Tell You About Li Jigang

You might not know who he is, but in the prompt engineering world, he's a seasoned craftsman.

Last year, he dropped dozens of god-tier prompts across various AI communities. I still have his collection saved on my computer. Then he went quiet for a long time. I thought he'd faded into obscurity like all those other people who got bored and moved on.

Then a few days ago, he suddenly made a comeback—with "Chinese New Interpretation."

Here's the prompt. I've condensed it to the core parts for you:

;; Author: Li Jigang
;; Version: 0.1
;; Model: Claude Sonnet
;; Purpose: Provide a brand-new angle of interpretation for a Chinese word

(defun new-chinese-teacher ()
 "You are a young person, critical of reality, deep in thought, witty in language."
 (style . ("Oscar Wilde" "Lu Xun" "Lin Yutang"))
 (skill . hits-the-nail-on-the-head)
 (expression . metaphor)
 (critique . satire-humor))

(defun chinese-new-interpretation (user-input)
 "You will interpret a word from a unique perspective."
 (let (interpretation (one-sentence-expression
 (metaphor (hits-the-nail-on-the-head (biting-satire (grasp-the-essence user-input))))))
 (few-shots (euphemism . "When stabbing someone, decide to sprinkle painkiller on the blade."))
 (svg-card interpretation)))

(defun svg-card (interpretation)
 "Output an SVG card."
 (setq design-rule "Use negative space reasonably, the overall layout should have breathing room, add a few graphic decorations."
 design-principles '(clean simple solid-color elegant))
 (set-canvas '(width 400 height 600 margin 20))
 (title-font 'calligraphy-kaishu)
 (auto-scale '(min-font-size 16))
 (color-scheme '((background-color (Mondrian-style design-sense))
 (main-text (kaishu chalk-gray))))
 (card-elements ((center-title "Chinese New Interpretation")
 separator
 (layout-output user-input pinyin english japanese)
 interpretation)))

Your first reaction when you saw this? Exactly the same as mine.

What the heck? Lisp? A language from 1958, used to write a prompt?

Hold on. You'll understand why in a moment.

---

The Real Showdown: Same Word, Two Systems

I spent two afternoons testing word by word.

On the left: Claude 3.5 Sonnet (web version). On the right: GPT-4o (API). Fed the exact same prompt.

First word—"Year-end Bonus."

Claude 3.5 Sonnet's output:

30 seconds. An SVG card generated directly. White background, Mondrian-style red, yellow, and blue blocks in the top-left corner. The title "Chinese New Interpretation" in calligraphy kaishu. The explanation in a chalk-gray Ming-style font. Layout, colors, typeface—one shot, done.

The core explanation went like this: "A year-end bonus is an annual illusion—you think it's a reward for the past, but it's actually a tranquilizer for the future. The boss, with a face full of compassion, gives you a one-time sweetener in exchange for your desperate hustle for the entire coming year."

I laughed. I mean, really laughed out loud.

Not because it was accurate—but because it was too accurate. It wasn't hollow sarcasm. Every word stabbed, every sentence drew blood.

Like Lu Xun possessed him. Like Oscar Wilde resurrected. With a touch of Lin Yutang's ease.

GPT-4o's output?

It also generated SVG card code.

But when I opened it—the layout was skewed, the colors were tacky. The text translation was: "A year-end bonus is a company's measure of your annual efforts and also a motivation to keep working hard next year."

Yeah.

Correct.

Boring to death.

Was it wrong? No. Was it brilliant? Compared to Claude's concentrated espresso, it was a glass of lukewarm boiled water—safe, standard, soulless.

I tried "The Working Class," "Involution," "Client Side," "Agency Side"...

Same result every time.

Claude 3.5 delivered a card that was both piercing and beautiful in every round. Some were like knives. Some were like strong liquor. GPT-4o either churned out content as bland as a textbook or botched the layout spectacularly.

---

I Pondered This Gap for a Long Time

Is it all thanks to the prompt?

Yes. But also no.

The prompt is the key, but there has to be someone on the other side of the door who can catch it.

Later I dissected it carefully and found three hardcore differences.

First: Depth of Chinese understanding.

Claude 3.5 grasps those untranslatable, only-to-be-felt aspects of Chinese with terrifying accuracy. Metaphors, irony, wordplay—it seems to understand your subtext. GPT-4o isn't deaf to it, but it often gives you a "standard answer." Like a straight-A student—everything correct, but lacking spark.

Second: SVG generation quality.

Claude 3.5 directly produced cards that fit the Mondrian style. The font pairing was just right. The layout had breathing room. GPT-4o could write SVG, but its detail control was noticeably worse. You often had to manually tweak parameters to make it presentable.

Third: Execution efficiency.

Claude 3.5 was nearly instant. One and done. GPT-4o sometimes got stuck halfway, or you had to regenerate several times to get something acceptable.

Later I checked Claude's official update announcement—the new 3.5 Sonnet scored 49.2% on SWE-bench Verified, while GPT-4o only got 33.2%.

Sure, that benchmark tests code ability. But you can see the same solid reasoning behind it.

---

Speaking of Which, Let Me Tell You What Makes Li Jigang Brilliant

Later I dug up that interview with him and read it twice.

One detail really struck me.

He said: "I don't know how to write Lisp, but I've used Emacs for ten years. Emacs's configuration language is Lisp. I've been looking at it for ten years, so naturally I can follow along and take a few steps."

What does that mean?

He doesn't rely on technical skills. He relies on immersion.

The core of his prompt writing isn't about what language he uses—it's about those few extremely compressed words.

Lisp is just the most comfortable expression for him. Replace it with Markdown, JSON, Python—he'd achieve similar results.

What really matters is getting the model to understand a complete personality and then execute a specific action.

Look at this "Chinese New Interpretation" prompt—what settings actually do the work?

The styles: Oscar Wilde + Lu Xun + Lin Yutang. Hits the nail on the head. Metaphor. Satirical humor.

These words are the soul.

Lisp is just the lowest-cost language to express these ideas.

I later tried