When to open: You repeatedly saw the same type of questions in the first few months on newtype Planet (how beginners get started, what tools to use, network configuration...), and wanted a collection of answers to "frequently asked high-frequency questions". This wiki is a curated selection of Q&As from 2025-01 to 2025-07 compiled by newtype himself.
Difference from newtype Mindset · Super Individual Methodology the difference between: That one is a mindset system, while this wiki is specific answers to specific questions — practical issues such as tools, networks, platforms, subscriptions, Agents, etc.
1. Beginner Learning Path
1.1 Where to start with zero foundation
"Learning by doing, learn while using."
Steps:
- ChatGPT, Gemini, Claude Subscribe to all three / try all three
- Try Deep Research on all three, compare results
- Artifacts / Canvas / canvas (mainstream feature)
- Models at different tiers (fast, cost-effective vs strong reasoning, high-cost)
- Upload extra-long PDFs and let Gemini and ChatGPT do full-text translation, compare differences
- After getting familiar with them → use ChatWise / Cherry Studio(AI client)
- to see the difference between first-party and third-party implementations
- Get exposed to API concepts,OpenRouter platforms, etc.
- "Leaving the newbie village" milestone: understand and be able to get started
1.2 How to improve technical cognition without a programming background
"You don't have to be good at writing code, but you must be able to understand it."
newtype's own path a year and a half ago (starting from zero):
- Learn some Python basics
- Learn LangChain, write scripts hands-on, build the simplest Chatbot, handle historical conversations with limited context windows
- Learn RAG, practice hands-on, understand the performance differences of various RAG strategies
- Learn Agent frameworks(AutoGen / CrewAI), hand-build workflows
- Force yourself to read AI papers → now his judgments on the industry and VCs are basically consistent
Conclusion:Learning by doing, don't hesitate, just do it.
1.3 Prompt engineering doesn't require taking dedicated courses
"The essence of prompts is logic and methodology."
- How to onboard new employees at a company? Give goals / background / path / reference cases — that's the prompt structure, nothing mysterious
- Learning logic / books like the Pyramid Principle >> taking prompt courses
- newtype doesn't watch any bloggers or content that talks about prompts
2. Network Configuration and Product Selection
2.1 Network conditions are a prerequisite
"If you can't get past the network hurdle, you definitely can't use AI."
| Product | Network requirements |
|---|---|
| Claude / Claude Code | Clean US residential IP(without it, don't even bother) |
| NotebookLM | Open in some regions (US market) |
| Gemini | Network lenient, China-friendly |
| Grok | Network lenient |
| ChatGPT | Quietly dumbed down (domestic IP) |
2.2 Claude Code subscription vs API
"Using Claude Code via API is extremely, extremely expensive, you can't skimp if you want good results. The most cost-effective way is to subscribe — a $20 USD Claude membership lets you use Claude Sonnet 4, and after you use up your quota you wait a few hours for it to reset. It's enough for daily use."
2.3 Three reasons for not subscribing to ChatGPT (newtype's own)
- Quietly dumbed down(domestic IP); static IP can circumvent this, but he detests this practice
- The model is no longer leaps and bounds ahead — Claude / Gemini / Grok have all caught up
- Search is a weak point, information is not comprehensive enough (compared to Gemini)
Just call it via API when needed.
2.4 Recommended comprehensive tool stack (based on network conditions)
| Your network situation | Primary recommendation |
|---|---|
| Have a US residential static IP | Claude(strong logical ability, MCP support, top-tier programming) |
| Domestic network + Google accessible | Gemini(balanced / no IP requirement / 1 million context) |
| Have Twitter Blue verification | + Super Grok (auxiliary) |
| Universal applicable | OpenRouter(aggregates all models + ChatWise / Cherry Studio client) |
2.5 Top-tier products are the most cost-effective
"As long as you choose top-tier products, they are definitely the most cost-effective."
Domestic platforms are a tier below overseas platforms. Don't pick through garbage in a garbage dump.
3. Multi-platform Strategy
3.1 newtype's own platform distribution
| Domestic | Follower count | Overseas | Status |
|---|---|---|---|
| WeChat Channels | 68,000 | YouTube | Long-term main front |
| Bilibili | 40,000 | Medium | Long-term |
| Douyin | 40,000 | X (Twitter) | Long-term |
| Xiaohongshu | 10,000 | Just started | |
| WeChat Official Account | (main text front) | newtype.pro | (Google SEO) |
Long-term strategy:YouTube + X + Medium + Reddit + newtype.pro operating five platforms in parallel. Domestic platforms are "a garbage dump, not worth it" (slightly emotional but reflects his priorities).
3.2 3 reasons for not joining or creating WeChat groups
- Time and energy consumption is too high — you have to charge tens of thousands in ticket fees → users expect unlimited "private board" treatment → you will inevitably be scolded for "scamming leeks" → the model is unsustainable
- Early stage of AI development = filtering > education: People who aren't even willing to proactively Google really can't learn this
- Don't provide emotional value, adhere to Intellectual Honesty
4 community platforms: Knowledge Planet (domestic, most content, strongest interaction), WeChat Channels membership, YouTube membership channel, Patreon.
4. AI Coding Tool Stack
4.1 Cursor vs Claude Code
"Cursor is a 'luxury sports car' with strong current performance and smooth experience; Claude Code is a 'concept car' pointing to the future. "
| Dimension | Cursor | Claude Code |
|---|---|---|
| Positioning | AI programming assistant | Intelligent programming partner (understands high-level intent + autonomous tasks) |
| Style | Passive assistance | Active autonomy |
| Depth of task automation | Medium | High (aggressive) |
| Current maturity | High | Medium |
| Long-term potential | Medium | High |
Current best practice:Use the two together, switch back and forth. Claude Code can only run after configuring a static IP (network issue).
4.2 Windsurf is out
- OpenAI is going to acquire Windsurf → Anthropic cuts off Claude 4 supply → Windsurf loses its top-tier model → no need to consider
- OpenAI is clearly behind in AI coding, no need to waste time on Codex
4.3 Why AI Coding is important (underlying reason)
"Code is the embodiment of pure logic and structured thinking — strict causality, step planning, complex conditional judgments. This is completely different from the ambiguity of generating love poems or emails.AI programming is the ultimate training for a model's logical reasoning and planning abilities, and is also related to the cultivation of long-context understanding and precision, and systematic 'tool use' capabilities."
"So,companies and models that can't do AI coding well aren't worth looking at, they have no future. "
5. newtype's Current Tool Stack
5.1 Daily main use (2025 H1)
对话 主力: Grok 3 + Gemini 2.0
辅助: ChatWise + OpenRouter + Ollama
搜索 主力: Perplexity
辅助: Grok 3 + Gemini 2.0
文章转存 Obsidian Web Clipper
知识库 Obsidian + Copilot
编程 Cursor + Claude Sonnet
5.2 ChatWise configuration
4 permanent MCPs:
- filesystem(local files)
- tavily(search)
- sequential-thinking(reasoning chain)
- markitdown(file to markdown)
2 functions of ChatWise:
- Backup: when other AIs are down or hit quota limits (had 3 Claude accounts banned), you can use it
- Early access: try new models / new MCPs first on ChatWise, compare how different models call the same MCP
5.3 Hidden uses of Cursor
"Cursor is AI programming software. But due to its native capabilities, it is also suitable for knowledge work."
- Use it as an AI knowledge base — open an ordinary folder storing Markdown files, it will create an index, supports all top-tier models
- All top-tier models supported
- Use Cursor for work involving local knowledge bases
5.4 Introduction to key software
| Software | Purpose |
|---|---|
| Milvus | Local vector database |
| OrbStack | Docker for Mac, used to run these projects |
| Raycast | Mac efficiency hub, $8/month, AI Chat summoned via shortcut key |
| GPUstack | Multi-machine parallel computing coordination (fine-tuning) |
| Llama Factory | Fine-tuning tool |
6. Prompt House Design Philosophy
6.1 Why build Prompt House
"Context Matters! The Prompt House and Obsidian MCP I've already released, as well as the upcoming plugins, are all based on this concept.Products are carriers used to express ideas. "
6.2 2-step MCP Tool call design
Problem: OpenMemory MCP stores all content (prompts + miscellaneous stuff) together, leading to low LLM filtering rates.
Prompt House's solution:
- Step 1: The LLM first retrieves the Prompt list(titles + tags, no content) → less information, fast filtering
- Step 2: After the LLM makes a selection, it then returns the specific content
Key optimizations:
- Accurate titles and tags → improve LLM accuracy
- Clear structure, small amount of text → improve LLM efficiency (filtering speed)
- Add an "AI polish" feature — users don't have to fill in titles and tags, AI fills them in
6.3 PS: Pipeline Prompt calls
A YouTube comment pointed out: "Through Prompt House's MCP, you can let the LLM complete the entire Pipeline of Prompt calls. If designed well, it's a very lightweight Workflow."This counts as an advanced usage. "
7. Bearish on Workflow, Bullish on MCP
7.1 Bearish on Workflow orchestration
"I'm not bullish on Workflow orchestration:
- Very few people have the ability to orchestrate — you need to understand both business and AI
- This approach imposes heavy restrictions on AI
- Orchestration level is the ceiling"
7.2 Bullish on MCP + Agent autonomy
"Let AI autonomously call tools and decide how to execute. That is the most correct direction."
Future architecture (analogous to an octopus): There is one core brain, and each tentacle also has a secondary cerebellum. These brains/models can all autonomously call a number of tools via MCP.
7.3 The essence of MCP
"MCP = USB-C. Its value depends on how many devices adopt the USB-C interface."
| Dimension | API | MCP |
|---|---|---|
| Format | Fixed | Dynamic |
| Semantic transmission | Not supported | Supported |
| Flexibility | Low | High |
| Connect 1 piece of software, get it working | Large workload | Equivalent cost |
| Scale to 100 pieces of software | 100x workload | Workload basically unchanged |
7.4 Two main lines of AI development
- Master more information
- Control more tools — MCP is the key to this main line
8. Video Production Workflow (A Roll + B Roll)
8.1 A Roll (on-camera footage)
- Write script → record with camera, set up teleprompter → read along (requires a lot of practice)
- Key tip: While writing the script, simulate the video frames in your mind — what expression, what B Roll for this segment
- A script is not simply an article
8.2 B Roll (supplementary / operation footage)
- Almost entirely screen recordings
- Use Screen Studio($ paid, Mac only) — the effect where the mouse hovers over and it automatically zooms in
- or OBS(free, used by content creators worldwide)
8.3 Narrating while operating (live-stream feel)
Add 2 sources in OBS:
- Audio: microphone + system sound
- Video: screen capture + webcam (overlaid, placed in the corner)
But newtype does not use webcam overlay (poor image quality). Instead: record screen + USB microphone on one side / film the person with a camera + shotgun mic on the other, sync audio track waveforms in post-production, and mute one segment of audio.
8.4 Core insights on video production
"Things related to content expression and creation are just as valuable as AI-generated content."
Nine, AI Writing & Style Definition
9.1 Claude is the top choice for AI writing
"In my personal experience with AI writing, Claude is the best. It has the strongest logic. I usually discuss topics with Claude during the topic selection and preparation stage. It always gives me something new. After I finish writing a video script, I also run it by Claude; out of the three revision suggestions it gives, I adopt at least one."
9.2 First use AI to distill style definitions
"99.9% of people don't know exactly what a so-called 'style' is. You can first give AI some materials, let AI perform 'style distillation' to get a very clear, detailed description, then use that description together with more materials to issue your request."
9.3 The core of AI writing
"When using AI to write, the most core thing is not the AI tool, but how much the person actually understands about writing. "
9.4 Writing copy is a fundamental skill
"Even if AI can do it, you still need to practice.You've succeeded when you practice to the point where you look down on what AI writes. This is your livelihood. The core of writing copy is logic. Don't get sidetracked thinking it's about words and phrases; those can be ignored. With good logic, you can write extremely awesome copy even in plain, everyday language."
Ten, Look Down on AI PPT + Replace It
10.1 No products that meet the standard
"Currently, all AI PPT tools fail to reach my 60-point passing standard."
It's not that I'm not optimistic about AI,it's that I'm not optimistic about PPT — through code, AI can create various presentation formats such as HTML that are more beautiful, more intuitive, higher in information density, and more interactive.I don't look at things that are destined to become obsolete at all.
10.2 Recommend learning Keynote
Xu Cen's Keynote course, which I learned ten years ago,has benefited me to this day.
10.3 Breakdown of agent writing copy + PPT layout (in response to a user's question)
"I suggest you split the plan into two parts:
- Text logic(text on each PPT page) — first write it well in Word,call the model via API + use prompts
- Visual logic(PPT layout) — PowerPoint supports VBA / Office JS API code control (JS API is better)
Integrate these two parts into a Workflow, package it as an agent solution, and you're basically done."
Eleven, RAG vs Fine-tune
11.1 Applicable scenarios for Fine-tune
- Professional domain knowledge(medicine / law / finance) — mastering professional terminology and background
- Task specificity(sentiment analysis / text classification)
- Style and tone adjustment
11.2 Applicable scenarios for RAG
- Knowledge-intensive(customer service) — retrieving external information
- Dynamic knowledge updates(news) — real-time retrieval
- Interpretability and traceability(law / healthcare) — answers can be traced back to sources
11.3 RAG vs Fine-tune decision
"If you can't figure out RAG, I don't recommend jumping into the big pit of fine-tuning."
Key insights(regarding the poor RAG performance of 70B local models):
- There is a world of difference in the comprehensive capabilities of models: The training/inference resources for Qwen Max far exceed those of locally quantized versions; it is stronger in instruction following, handling contextual nuances, integrating multiple retrieved fragments, and tolerating imperfect context quality
- Different minimum requirements for retrieval quality: Strong models have high tolerance for current retrieval strategies, while weak models require more precise retrieval
Optimization directions: Replace the Embedding model / adjust chunk size and overlap / design prompt templates.
Twelve, Note-taking System (Input-Processing-Output)
12.1 Reasons for choosing Obsidian
"Obsidian = A private and flexible writing vault that thinks."
- Local First: Notes are stored as Markdown files in the user's local folder
- Markdown ensures long-term readability and portability(even if you stop using Obsidian in the future, your notes will still be accessible)
- Multi-device sync relies on iCloud(iOS / macOS / Windows)
- This is the core reason for newtype's choice
12.2 Differences from Notion
"Notion is an All-in-One Workspace, an integrated workspace.It aims for full integration, with itself as the core. Some people like that, but it's not what I want." "I don't accept Notion-like tools; they can only be stored on their servers, which carries the hidden risk of being held hostage. "
12.3 Obsidian file structure
newtype's Vault structure:Input → Processing → Output
- Input: various articles, papers, etc. that you want to read carefully (top of the funnel)
- Processing: domain-specific accumulation (being organized)
- Output: final deliverables (for example, scripts for all videos are stored here)
12.4 Batch converting files to Markdown
- Released by Microsoft markitdown(GitHub)
- MinerU(also open source)
- Paid product (Tencent Cloud Large Model Knowledge Engine)
12.5 Don't deploy locally unless absolutely necessary
"If you really don't feel comfortable storing your materials on others' servers, you can use Milvus to build a local vector database, then connect it to various AI tools via MCP. This is the most flexible approach I can think of that also meets local storage needs."
Thirteen, Don't Rely on Willpower, Three Things
"The saying 'self-discipline determines freedom' is harmful. Because when 'self-discipline' comes into play, you are already not free, and there will be a lot of internal friction. Therefore no need for self-discipline. My weekly video updates and daily content posts in the community all happen naturally."
13.1 Three things (to replace "self-discipline")
- Sense of mission — is the source motivation.The level of sense of mission is the biggest difference between people. newtype's sense of mission is creation and pioneering
- Physical strength — determines execution. Exercise every day, eat scientifically, get enough sleep
- Don't force it — the prerequisite for no internal friction
13.2 Psychological counseling (seeking inward)
- newtype himself is undergoing long-term counseling, at a fixed time every Saturday afternoon
- Platform:Jiandan Xinli (Simple Psychology)
- Chose a young counselor (since long-term counseling is needed, counselors in their fifties or sixties have a higher chance of passing away first)
- Compares himself to the Coach role of Wendy in *Billions* = helping employees get back to their peak state
- "The four elements of spiritual practice: wealth, companions, method, and place. The counselor is that 'companion'"
13.3 Death anxiety
"The direct cause is death anxiety. There are many reasons behind it — not just age stage or family of origin. Starting in 2010, I actively discarded all the things my family and society taught me in the past, and re-cultivated myself from scratch."
Fourteen, Mobile AI Is Not Yet in Place
"On-device AI, especially on mobile, has significantly fallen behind the industry. "
Two key limitations:
- Permissions: Without system-level permissions, many things can't be done. The permissions are held by phone manufacturers and won't be released
- Computing power: The mainstream model on phones is 3B, and general-purpose chips seem to support up to 7B. Such small models definitely require targeted fine-tuning
Advice for developers: Don't focus too much on AI at this stage,return to application thinking. Look at how AI can add points on the premise of making the application well. The underlying infrastructure is still a mess with too many restrictions.
Fifteen, Views on Agent
15.1 Don't look at Agent frameworks, look at actual projects
"I don't look at these Agent architectures anymore.The focus at this stage is not what the architecture is like, but what you want to build. There are already many ready-made open-source projects that can be used directly. If I were to build a product with an Agent, I would directly take apart Gemini CLI and use it as the core."
15.2 The only mature Agent application at present
"At present, the only mature Agent application is Deep Research. If an individual wants to build something, use MCP."
15.3 Ranking of Deep Research capabilities (newtype's self-assessment)
OpenAI > 类实习生 > Gemini > Perplexity > Grok = 秘塔 > Deep-Research(最佳开源)
OpenAI's Deep Research can be sold for $200/month, it is truly "usable". Other products don't even reach 60% accuracy.
15.4 OpenAI's Deep Research is slow because it "piles on computing power"
newtype speculates: OpenAI's Deep Research is much slower than others because it is built by piling on computing power. Therefore it's not complicated; the question is why other players don't pile on computing power. Doing some cross-validation can improve it further,it's a process issue.
Sixteen, newtype's Personal Background
16.1 Not from a programming background
"My programming ability is very elementary."
16.2 Three core abilities (cultivated over 10 years in the communications and marketing industry)
- Information gathering and rapid learning(essential for pitching in completely unfamiliar industries)
- Strategic ability:Defining problems and finding solutions
- Narrative ability(content ability, persuasion ability, sensitivity to words and emotions)
16.3 New skills gained from a short entrepreneurial experience
- Ability to build products
- Not bound by professional background: If I had no UI experience, I would look at Dribbble every day to learn/copy from others; if I didn't know professional software, I would use Keynote to draw UIs
- Advance by any means necessary
16.4 Inner core + external flywheel
"These are just external abilities. I think my strongest point is my inner core, a personality that is gradually being improved and becoming complete. This is the result of more than ten years of continuous [seeking inward]. The inner core plus external abilities form two flywheels. Once they start spinning, everything you want will come along with it, without extra effort.JUST DO IT. "
Seventeen, Regarding the Company Strategy "All in AI"
17.1 The root cause why the boss makes the decision but can't get it done
"At the company I work for, the boss proposed the All in AI strategy. But after two years, nothing has been achieved.The root cause is that the person making the decision doesn't understand it at all, only stays on the surface, only talks about empty buzzwords, and really enjoys the thrill of formulating strategy. "
17.2 Deep learning does not equal becoming a professional
"My expressions of views on AI are completely not in the tone of an industry insider, but the core meaning I express is the same as theirs,and I can even explain it more clearly than them. "
"So-called 'professionalism' is still not deep enough.In 99.9% of cases, 'professional' people only know more than outsiders because they are in the industry. It doesn't count as specialized/deep."
17.3 AI can only assist
"AI can only assist. It's already very good if it can give some inspiration. "
Eighteen, Regarding Deep Learning vs Deep Research (Decision-making Positions)
User question: As a corporate executive, making judgments and decisions doesn't require professional-level understanding. Is Deep Research + Reasoning enough?
newtype's answer:It is necessary (to continue deep learning)
"For the vast majority of things, you think you've mastered them enough to make decisions, but you're actually far from it." "For example, the boss at my company proposed All in AI, but didn't understand it at all, only stayed on the surface → achieved nothing."
Nineteen, Regarding Buffett Investment Thinking Training
User asks: Can I feed all of Buffett's letters and interviews to AI to train it to have Buffett's thinking? Then throw company annual reports at it for analysis?
Answer:
- Buffett / Graham material models already exist (or can be searched),you don't need to provide them
- Can this lead to reliable investment conclusions?Various providers have Deep Research, give it a try
- With current technology, it can only give you some inspiration
- Investment thinking is only part of Buffett's success, others really can't replicate it
- If you really want to do it: use AI to have a conversation with your own methodology, see what new things it can give within your framework
Twenty, Overview of Q&A (20+ Q&As)
Core topics covered
| Topic | Number of Q&As (estimated) |
|---|---|
| Beginner onboarding / tool stack | 4-5 |
| Network configuration / subscription decisions | 3 |
| AI Coding / Cursor / Claude Code | 3 |
| Knowledge base / Obsidian / RAG | 6 |
| Prompt House / MCP design | 2 |
| Video production workflow | 1 |
| Platform strategy / traffic | 2 |
| Mindset / no self-discipline / psychological counseling | 3 |
| Agent / Workflow / frameworks | 4 |
| Fine-tuning vs RAG / deployment | 4 |
| Model selection / DeepSeek / Claude | 3 |
| Commercialization / Solo Entrepreneurship / Corporate Strategy | 4 |
21. Connection with Existing Wikis
- newtype Mindset · Super Individual Methodology — newtype Mindset System (this wiki is its concrete application in Q&A scenarios)
- Full Evolution of the newtype Knowledge System and Life OS — Full Evolution of the Knowledge System (Sections 11–12 of this wiki are a simplified version)
- newtype · AI Coding Practice and Toolstack Evolution — History of AI Coding Evolution (Section 4 of this wiki is its Q&A supplement)
- newtype · AI Toolstack Practice and Evolution — Toolstack Details (Section 5 of this wiki is its Q&A snapshot)
- newtype · Vibe Coding Independent Product Practice — History of Independently Developed Products (Section 10 of this wiki covers the Prompt House design philosophy)
- Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk) — Cal Rueb Claude Code Talk (Section 4 Q&A of this wiki is complementary to it)
History
- 2026-05-17:Phase 4 ingest from
13_newtype星球问答_纯享版.pdf(Q&A from 2025-01 to 2025-07) created. Fully covers 20+ high-frequency Q&A topics: newcomer paths, network configuration, platform strategy, AI Coding, Prompt House design, video production, RAG vs Fine-tune, mobile judgment, views on Agents, core + external flywheel, etc.