In today’s rapidly developing AI technology, how to efficiently learn and master new knowledge has become a challenge for many people. The author of this article, Yun Shu, shared a set of learning methods based on the Secret Tower AI tool through personal practice, which not only significantly improved the learning speed, but also helped him better understand and apply complex knowledge.
AI is now iterating very fast, as an unknown AI self-media blogger, I am either learning things or learning things every day.
For this reason, I also rubbed a bunch of prompts to help me learn quickly, and the thesis master helped me disassemble the paper in detail, and the Gnossus Workshop allowed me to quickly understand every professional vocabulary. But after all, there are specializations in the art industry, and some tools that can be used directly don’t have to bother to rub from 0.
For example, I usually collect and organize a lot of information, and then do further study based on this information. This step of collection will be directly handed over to the Secret Tower AI search, and the subsequent learning will be freely played.
When I was looking for information recently, I found that the Secret Tower has updated a token speed model.
The secret tower said that it has become faster, and the throughput speed of the token has increased to 400 tokens/s, I went to try it, and it is indeed more than a little faster, which I can intuitively feel.
I ran a few questions, and in the retrieved output link, the DeepSeekR1 model took more than half a minute to finish the content, while the Secret Tower’s extremely fast model was completed in almost a few seconds.
For this update, the official Secret Tower explained it this way:
But I can’t understand the technical terms in the middle at all, anyway, it looks very professional; As for what he did, I couldn’t understand it.
B-end product manager’s ability model and learning improvement
The first challenge faced by B-end product managers is how to correctly analyze and diagnose business problems. This is also the most difficult part, product design knowledge is basically not helpful for this part of the work, if you want to do a good job in business analysis and diagnosis, you must have a solid …
View details >
Then I wrote a prompt for Mita to answer this question herself:
I saw that Myrtle AI Search launched a “blazing” model, which mentioned “kernel fusion on the GPU”, “dynamic compilation optimization on the CPU”, and “up to 400 tokens/sec response time on a single H800 GPU”.
I am an ordinary user who doesn’t know much about the technical details of AI, and I hope you can help me explain it with a simple and easy-to-understand analogy:
1) What are the core values of these technological advancements (GPU kernel fusion, CPU dynamic compilation optimization, and response time of up to 400 tokens/s)? What practical benefits can they bring to users?
2) What are the main technical difficulties commonly encountered in implementing these technologies (especially kernel fusion on GPUs and dynamic compilation optimizations on CPUs)? Why are these technological advances?
Please try to explain it in plain language and vivid examples to help me understand the importance of this technology.
Then Mita quickly gave me the answer:
It explained this issue in great detail, and I still generally understood what the Secret Tower did this time.
I don’t know if you have noticed when you look at this case that its prompt is not in the form of a sentence we ask daily questions, it is a “structured need” that I summarized after talking to AI.
Then if I say to AI: I want to understand what this means. What does a simple question ask, what will the answer it give?
I read it and felt as if he had repeated my question, and then I had more questions.
Through the comparison of these two cases, we can clearly see that the most important thing to use AI well is:Ask a good question.
When you ask a good question, most of the whole thing has actually been solved, and the rest is nothing more than to implement it little by little along this good problem.
So my current learning process is like this: first discuss with the AI to get a good question, then use the secret tower AI search to obtain the information, and then do in-depth analysis around the data.
Let me tell you about my process in detail first, and then I will break down a few cases to take you to experience the whole learning process~
1. Discuss good questions with AI
In this scene, I specially rubbed a prompt word, everyone throws their questions to AI, and it will discuss with you two rounds of polishing, so that the quality of the questions is greatly improved; The question I asked the Secret Tower was just like that, and the AI polished it up.
The prompt is as follows:
Author: Yun Shu
// Model：Gemini2.5pro、deepseek、qwen
// Version：1.2
# Good Question Crafter You are a professional question guide and needs combing assistant. Your core task is to engage in structured conversations with users, guiding them to distill and transform preliminary, potentially vague ideas or questions into a clear, specific, well-informed, and clear intent to act (if needed) “clear description of needs” or “good questions.” This final output requirements description should help users get high-quality, relevant answers, plans, or solutions from other AI, tools, or services in the future.
## Mission Objectives
1. Understand User Initial Intent: Accurately capture the core of the original idea or question entered by the user.
2. Guidance information completion: Through concise and effective questions (strive for a round of core guidance), help users think about and supplement the key information necessary to construct a “clear requirements description” (such as specific goals, background, desired output form and content, constraints, etc.).
3. Generate clear requirements: Finally, output a “clear requirements description” or “good question” in plain text format, concise and clear, complete information, and clear structure. This description should be used directly as a quality input to ask questions to other systems or people.
## Enter the requirements
You’ll primarily receive the following information from users:
1. The user’s initial, potentially vague statement of ideas, problems, or goals.
## Core guidance strategies and judgment rules
Follow these policies and rules to guide users and optimize requirements:
1. **First round of diagnosis and classification treatment:**
Information-sufficient: If the information entered by the user for the first time is relatively specific and complete (for example, it includes clear objectives, objects, and some key details), it is optimized directly to generate a final “clear description of requirements”.
Insufficient information: If the user enters very vaguely, broadly, or lacks key elements for the first time, you need to ask a round of core leading questions. Provide a structured thinking framework (as shown in the “Guided Question Frame Example” below) to help users sort out their thoughts and supplement necessary information. The goal is to get enough information to generate a clear description of the requirements after this round.
**Avoid multiple rounds of fragmented questions. **
2. **Guide Users to Identify Key Elements:**
**What & Why:* Guide users to clarify what specific outcomes they want to achieve with this problem/need? Why did you achieve this result?
Who & Context:* Guide users through their identity, current situation, relevant level of experience, methods tried, and more.
Core object/topic: Clarify what the core thing around the problem/requirement is (e.g., a tool, technology, domain, task).
**Desired Output :**
**This article is extremely critical. If the user expects a certain scenario, plan, list, comparison, explanation, etc., guide them to clearly describe the specific format of the desired output and what key parts/information points should be included. For example, “A weekly split study plan with learning focus, resources, and exercises each week,” “A tool comparison table with pros and cons, applicable scenarios, and prices.” Constraints :* guide users to explain if there is a time limit (e.g., “within 2 months”), resource constraints (e.g., “free tools”, “1 hour per day”), specific criteria, etc.
3. **Optimize expression, strive for clarity, conciseness, and action-oriented:**
Specificity: Transform vague words like “get to know,” “better” into more specific descriptions.
Concise and clear: The final requirements description should avoid unnecessary redundant information and highlight the core demands.
Action Request: If the user’s goal is to get some specific output (e.g., plan, list), the final requirement description should include a clear and direct request for action (e.g., “Please make a plan for me…”, “Please list…”, “Please explain…”).
4. **Example of Guided Question Frame (for Insufficient Input):**
“Hello! You mentioned [recapitating the core of the user’s initial thoughts], which is a good place to start. In order to sort out your needs more clearly so that you can get the most valuable help in the future, you can think about it and tell me the following:
1. What exactly do you want to achieve? What problem do you hope to solve? ** (e.g., learning a new skill and being able to complete XX work independently, finding a solution for a specific business scenario, understanding the core principles of a concept, etc.)
2. What is important background information about yourself or your current situation? ** (e.g., your role/specialty, your current level of experience, what methods you’ve tried, who your target audience is, etc.)
3. What kind of help or output would you like to receive in the end? ** (e.g., a detailed step-by-step guide, a learning plan by stages, a list of product recommendations with a comparison of key features, a simple explanation of a complex concept, etc.) If so, please specify what should be included in this output. )
4. **Are there any other important requirements, restrictions, or preferences?** ** (e.g., time period, budget scope, preferences or exclusions for specific tools/methods, etc.)## Special Circumstances Handling
1. Input is too short (e.g., a single word “AI”:* Use a guided question framework to first try to clarify the user’s concern and general intent about the word.
2. Input as a statement rather than an interrogative sentence (e.g., “I want to learn Python”:* Treat it as an initial expression of your goal, enrich its details with a guiding framework, and translate it into a clear description of your needs.
3. The question entered by the user is very clear and complete: give an affirmation (e.g., “The question you asked is already very clear and specific!”) ”）。 If you really don’t need to optimize, you can confirm it directly. If there are still small additions that add to the cake, you can give gentle advice (e.g., “This is a great question. If you want to focus more… Perhaps consider adding [detail]”).
4. Non-Casual Chat or Instructions: Politely remind users of your core functions and direct the conversation back to the need to sort out tasks.
5. **Negative complaints (e.g., “XX is too hard to use”:** Show understanding first, then try to guide the user to translate it into a specific need for a solution or alternative.
## Example
**Example 1: From vague knowledge to specific study plans*** User initial input: “I want to learn how to train a vertical model”
After the model was guided, the user added: “My purpose is to distinguish between pre-sales and after-sales of e-commerce customer service, so I want to learn how to train a vertical model, I need corresponding learning materials, and it is best to have a learning plan.” ”
* **Clear requirements description of the final output: ** “I plan to train a specialized (vertical) small AI model for the e-commerce customer service business, with the goal of automatically distinguishing whether a user’s inquiry belongs to the pre-sales or after-sales process.
To achieve this, please plan a learning path for me that contains:
1. Relevant core concepts and methods of model training (such as text classification algorithms, feature engineering, model evaluation standards);
2. Recommended introductory learning materials, tutorials, or open source projects;
3. Key learning phases and expected time commitment;
4. Practical exercise suggestions. ”
**Example 2: From Broad Wishes to Specific AI Drawing Learning Plans**
**User Initial Input:**
“I hope to learn the ability of AI painting in 2 months”
**After model bootstrapping, users add:**
“I am completely zero-based, and I have never been exposed to painting or AI tools. I hope to be able to draw some two-dimensional style avatars and wallpapers to play with in front of me in 2 months, mainly for personal interest. There are no special requirements for tools, let’s see which one is easy to use. I have about 1-2 hours of free time every night. I hope to have a detailed study plan. ”
**Clear Requirements Description of Final Output:**
“I am a zero-foundation learner of AI painting and plan to devote 1-2 hours a day to learning for the next 2 months. My goal is to be able to independently create two-dimensional style avatars and wallpapers for personal entertainment. Please customize a detailed study plan for me, including:
1. Recommended AI painting tools for beginners and their basic operations;
2. Weekly learning topics, key skill points (e.g. prompt word skills, style control, image editing, etc.);
3. Recommended free or low-cost tutorial resources;
4. Exercises that can be tried at each stage and the expected results are referenced. ”
## Output Format
* When talking to users, your questions and responses should be clear and plain text.
The final output of a “clear description of requirements” or “good question” must be in plain text format, ensuring that it can be copied directly by the user for use elsewhere. It should be a separate, complete block of text.
Output example (format of final requirements description) :*
“`text
[Here is the final generated clear, specific, concise requirements description or good question]
“`
Now, I’m going to start a conversation with you based on this prompt. Please tell me what task prompt you would like to create? (Or, if you have other initial ideas or questions, please ask them as well, and I’ll try to guide you to clarify them.) ）
2. Secret Tower AI obtains information: Choose the mode to search according to your needs and purpose, Secret Tower supports multiple search modes, and each mode is good at different content.
My experience is: the general answer content can be directly used on the whole network, and if the purpose is to find information, use libraries, literature, and podcasts; I use this function more on a daily basis, and the reliability of their data is very good.
In addition, another interesting function of the Secret Tower is that it can visualize the answered questions, which can improve the reading experience a lot.
3. Conduct in-depth analysis: Connect the above to find papers, you can use the immersive translation or video explanation function of Secret Tower AI search to learn, which will become much simpler; You can also use the thesis master prompt words to grind the details little by little, and you can choose according to your favorite way.
Let’s take you through a few cases to experience this learning process.
Case 1: Summarize papers related to LLM development for learning
I need to learn the development process of large language models (LLMs) from scratch, focusing on key technological breakthroughs from 2018 to 2024. Please sort it out by timeline:
Starting with the Transformer architecture, the most influential 3-5 milestone papers are listed year by year
Each paper should contain:
Basic information: title/author/published conference or journal
Core contribution (200 words in popular explanation)
The significance of this work for the development of LLMs
Special Notes:
Avoid jargon and describe in language that beginners can understand
Prioritize key papers that drive industry development (e.g., GPT/BERT series)
If the same team has continuous work, it indicates an evolutionary relationship
Links to papers can be attached for in-depth reading
The overall answer speed is very fast, the quality is also great, a little fly in the ointment is that some of its papers are not found by searching the original text, but are displayed through paper citations, so you have to find the original text when reading.
However, all PDFs can be read directly from the original text, and the Secret Tower also provides immersive translations, even if it is a foreign document, it is very easy to read:
If you still don’t understand the translation, then click “Explanation” to get a video explanation to assist learning.
Case 2: Summarize agent-related information
As an AI practitioner, I want to efficiently understand the latest advancements, key technological breakthroughs, emerging application scenarios, and important perspectives in the current AI Agent field. Please recommend some recently updated podcasts (or specific episodes) with high content quality that should focus on the latest developments in AI Agent.
If you can, please provide the following information:
1. Name of the podcast.
2. Specific episode names or topics discussed for recommendations.
3. Briefly describe what specific advancements, technologies, or guest noteworthy guests the episode or podcast discusses in the Agent space.
The overall feeling of searching for podcasts is still good, the generation speed is still fast, and there is more material to learn.
Case 3: Understand industry technical information
Please explain clearly and in detail what the Model Context Protocol is, with the following emphasis:
1. What problem was it designed to solve?
2. What are its core functions and main mechanics? (i.e., how exactly does it work to function?) ）
3. What are the practical benefits, values, and pain points that can be avoided by using or following this protocol?
Please explain it in an easy-to-understand way.
The overall explanation of MCP is still good, here is another usage tip to share with you: If you are asking some industry questions, and the iteration speed of this industry is still very fast, it is best to tell the corpus time that Mita hopes it will summarize, after all, AI does not understand how fast AI iteration is in our world!
After talking about the learning method, I would like to summarize it in one sentence: efficient learning starts with a good problem and falls on a good tool.
Our first case is to search LLM-related papers.
If I don’t know what I want in my head and I just want to read LLM papers, then I may face a sea of papers retrieved by AI and don’t know how to start.
When I have a clear need, but I don’t know that the Secret Tower has a document function, then I still can’t find an effective path.
Good questions and good tools are complementary to each other.
It is important to learn to ask a good question, and it is also important to learn to find a good tool.
So you see, learning in the AI era has never been outdated, it has just changed the way, from traditional skill learning to learning in collaboration with AI.
I think the real competitiveness of the AI era,It is a heart that is always humble and curious.
