A 17-year-old high school student wrote Thinking Claude, and netizens called it a divine prompt.
Recently, a 17-year-old high school student Tu Jinhao (@Richards Tu) developed a prompt word (Prompt) expansion tool called Thinking Claude, which sparked heated discussions in the Chinese AI discussion community, and some netizens called it "god-level" Prompt".
According to a report by (AI Post), this high school student is not only young, but he is also the champion of the AI track of Alibaba Global Mathematics Competition.
He observed that since the advent of OpenAI's ChatGPT-o1, AI's thinking chain has undergone major changes. It is no longer just a rigid reasoning process, but closer to the human stream of consciousness thinking mode. This observation inspired him to develop a prompt word tool for Claude 3.5 Sonnet with o1-like reasoning capabilities.
But it should be noted that Thinking Claude is not used to enhance the accuracy of the Claude 3.5 Sonnet model, but rather focuses on allowing the Claude 3.5 Sonnet model to demonstrate the reasoning process.
According to Thinking Claude's GitHub page, this project mainly consists of two core components: Thinking Protocol and Google extension.
Tu Jinhao stated that the Thinking Protocol is a comprehensive command set that guides Claude to engage in deep and systematic thinking before responding, and the browser extension is a tool that makes Claude's thinking process easier to read and manage.
Thinking Claude has the following 5 major features:
A consciousness stream that 'simulates' natural human thinking processes
The ability to continuously reflect and adjust through self-questioning and answering
The ability to accurately grasp user’s true intentions and needs
Logical coherence with a clear and traceable thinking process
The ability to self-adjust flexibly according to the scenario
How to use Thinking Claude? Step-by-step tutorial at a glance
How to use this 'Thinking Claude' in Claude? According to the instructions on the GitHub page, you need to first install the browser extension, which can be installed directly from the Chrome Web Store or manually downloaded and installed.
Click here to download the latest version of Thinking Claude compressed file
After installation is complete, users need to go to the Claude.ai website to perform 'Create & Edit Styles', and import the command set from the uncompressed file. Detailed instructions can be found in the following graphic explanation:
Source: Encrypted City charting How to use Thinking Claude? Step-by-step tutorial at a glance
Source: Encrypted City charting How to use Thinking Claude? Step-by-step tutorial at a glance
Source: Encrypted City charting How to use Thinking Claude? Step-by-step tutorial at a glance
Encrypted City tests Thinking Claude's logical reasoning ability
After successfully installing and importing the command set, you can start testing Thinking Claude's reasoning process!
(Encrypted City) provides it with some brain teasers or logic problems. Below are the questions and answer results:
Question 1: There are six cups in a row. The first three are milk, and the last three are empty. How can you rearrange the six cups by only moving one glass cup, so that the full cups and the empty cups alternate?
Source: webpage screenshot Encrypted City tests Thinking Claude's logical reasoning ability
Question two: A series of short brain teasers
Source: webpage screenshot Encrypted City tests Thinking Claude's logical reasoning ability
Source: webpage screenshot Encrypted City tests Thinking Claude's logical reasoning ability
Question three: 1=5, 2=10, 3=15, 4=20, what is 5=?
Source: webpage screenshot Encrypted City tests Thinking Claude's logical reasoning ability
From the above questions and answers, it can be seen that Thinking Claude demonstrates the process of 'simulating' human thinking logic. Some of the simulated processes are quite interesting, as if one can really imagine that humans think this way. Unfortunately, it failed on the third question, as the answer is 5=1.
This system allows Claude to display the reasoning process like the o1 model. Although it is interesting, YouTuber 'huangyihe', who focuses on the AI field, believes that the output patterns of large language models are fundamentally different from human thinking patterns, so we cannot say that Claude truly thinks. Thinking Claude is actually a smokescreen; it makes AI pretend to think and perform thinking. What we see as the thinking process is all generated scripts. However, having AI do inner monologues still has its 'performative value'.
'Thinking Claude Tutorial: A genius prompt written by a 17-year-old? Helps you test how smart it is.' This article was first published in 'Encrypted City'