newtype · AI Coding Practice and Tool Stack Evolution

Huang Yihe's AI Coding tool stack usage and mental models from 2024-09 to 2026-05. Covers all 315 posts in the ai-coding topic bucket. This wiki focuses on "how non-technical individuals can use AI Coding tools"(production-grade vibe coding in newtype · Vibe Coding Independent Product Practice).

When to open:

You want to get into AI Coding but don't know whether to choose Cursor / Claude Code / Codex

You use Cursor but your code always has bugs and want to know the workflow

You've heard of Augment / Trae / Windsurf but don't know if they're worth switching to

You want to understand why Claude Code suddenly became the main battlefield in late 2025

0. Timeline

graph LR
    A["2024-08~09<br/>Cursor 震撼<br/>(零代码做 Chrome 插件)"] --> B["2024-09<br/>v0+Claude+Cursor 三件套<br/>(newtype.pro 上线)"]
    B --> C["2024-10~12<br/>GitHub Copilot 跟进<br/>Cline 出现<br/>Cursor AI Review/bug finder"]
    C --> D["2025-02<br/>AI Coding 赛道全景<br/>DeepSeek/Claude 3.7"]
    D --> E["2025-03<br/>用 Cline 自建 MCP<br/>Obsidian MCP/Flux MCP"]
    E --> F["2025-04<br/>Gemini 2.5 Pro 大上下文扫码"]
    F --> G["2025-05~07<br/>Claude Code 走入主流<br/>Plan Mode / Ultra think 32K"]
    G --> H["2025-08~12<br/>Claude Code 主战场<br/>Sub-agents/Slash/Skills"]
    H --> I["2026-01+<br/>OpenCode/Codex/newtype-os<br/>多 Agent 编排"]

1. Lao Huang's jaw-dropping moment: Cursor (Post 471, 2024-08)

"The product that recently shocked the circle — Cursor. Before long,the most popular programming language in the world will be English. "

Shortly after in 2024-09 (Post 477):

"The shock Cursor gave me this time,is even stronger than ChatGPT two years ago. "

Key case (Post 473): Zero code + pure chat, had Cursor build a Chrome extension that summarizes webpages, done in 10 minutes (second attempt, with clearly written requirements).

Why the impact was so big:

Cursor is not an "AI assistant", it is a VS Code fork + deep AI integration
Natural language → code → debugging → deployment,the entire workflow in one tool
A "non-technical product manager" like Lao Huang could suddenly build complete products

2. AI Coding tool division of labor philosophy (classic Post 488)

v0 + Claude + Cursor three-piece suite(Lao Huang's distillation, 2024-09):

Tool	Role	Stage
v0.dev	UI / design mockups	Visual layer
Claude(Web/desktop)	Framework / architecture	Abstraction layer
Cursor	Details / debug / implementation	Implementation layer

Core mental model:

"Handing everything to one tool will most likely not work well. These AI tools are not yet strong enough for that,each has its own strengths. So accept this reality, stop yapping, JUST DO IT."

Anti-pattern: Trying to use Cursor alone to handle everything from UI to deployment → guaranteed to hit a wall.

2026 evolved version(Posts 2484, 2509):

Chief / Deputy → Claude (strong at planning logic)
Researcher → Gemini (large context)
Execution-type → Gemini Flash / Claude Sonnet (cheap)

See newtype · Agent Architecture and Multi-Agent Orchestration §8 Multi-Agent Orchestration for details.

3. AI Coding landscape overview (Post 838, 2025-02)

Lao Huang's curated AI Coding classification:

Category	Tools
Daily programming	Cursor / Windsurf / Trae
Web applications	bolt / v0 / lovable
Landing Page	Pagen / wegic
Frontend components	CopyWeb / screenshottocode
API	devv builder
Project prompts	copycoder
Mobile App	bolt / rork
Project documentation	codeguide

Lao Huang's personal choice(Post 556):

"If something is good enough, or I'm sure it's the best at the current stage, I stop looking at others. For AI Coding,I only look at Cursor and GitHub Copilot right now. "

Attitude toward domestic tools (same post): "I haven't touched domestic ones at all."

4. Comparison of code retrieval methods in AI programming tools (highlighted Post 1392, 2025-06)

4 retrieval modes:

Tool	Retrieval method	Advantages	Disadvantages
Cline	Real-time search (imitates humans, high-speed file content + regex)	Accurate, real-time	High token cost
Claude Code	Agentic Search (dynamically decides content/filename search + multi-round strategy)	Highest accuracy	Even higher token cost
Cursor	Classic Embeddings + RAG	High cost-effectiveness	Limited by Embedding comprehension ability
WindSurf	Same as Cursor but with low latency + enterprise deployment	Enterprise-grade	Slow initial indexing

Why it matters:

Cline's reason for rejecting RAG: A function definition is at line 200 / its call is at line 300 / its dependency is at line 400 → RAG can only read fragments, breaking the logic
Different scenarios suit different approaches; balance efficiency and quality

5. Cursor usage mental models

5.1 Three modes (Post 826)

Chat: Chat + manually select the code segment to apply → maximum user control
Composer: Edit multiple files → suitable for refactoring
Agent: Auto-execute → suitable for "I understand the business but don't want to write code"

5.2 Advanced Cursor recipes

Scenario	Recipe
LLM scanning codebase	Gemini 2.5 Pro (ultra-large context) → use chat to understand the entire project (Post 1071)
Use as a knowledge base	`.cursorrules` explicitly require "query all documents in the folder for every answer" + 4 MCP descriptions (Post 1154)
AI auto-writing notes	Composer tags documents with `[[]]` annotations + auto-creates empty.md files (highlighted Post 854)
Connect to Obsidian Vault	Cursor directly opens Obsidian Vault, folder Q&A (highlighted Post 483)

5.3 Cursor pitfall avoidance points (Posts 826, 1671)

Cursor modified previous modules → use Chat mode + manually select code / break the problem into smaller pieces
Code truncated due to exceeding token limit → mark [CONTINUE] to let the AI continue writing
5-module integration failure → incremental integration (2→3 modules, add gradually, check dependencies each time)

6. Claude Code main battlefield (2025-07~)

6.1 Why Claude Code became the main battlefield (Post 1586)

"Claude Code showed me the possibility of a relatively general-purpose Agent. "

Claude Code's 4 major stacks (Post 2148):

Stack	Function
Skills	Capability loading (encapsulate SOPs from your brain)
MCP	Tool calling (external server integration)
Slash commands	Workflow reuse (/research / /draft / custom)
Sub-agents	Multi-threaded division of labor

6.2 Ultra think 32K thinking budget (Post 1558)

Keyword	Thinking tokens
`Ultra think` / `ultrathink`	32K(HIGHEST)
`think hard` / `think harder`	10K(MIDDLE)
`think`	4K(BASIC)
(none)	0(NONE)

When to use: When Claude Code gives a wrong answer,first add "Ultra think" and retry; 90% of the time it's insufficient budget, not the model being bad.

6.3 Plan Mode (Post 1482)

Turn on Plan Mode, tell Claude the requirements + thinking (especially pain points and possible solutions)
It plans first,does not directly touch code
Suitable for complex tasks (such as Lao Huang's Obsidian MCP notes_insight new feature)

6.4 Let Claude Code maintain claude.md (Posts 1433, 2093)

Lao Huang's best practice:

"After finishing every issue,update CLAUDE.md once. If the issue is particularly large, when hitting certain milestones during the process, in addition to updating CLAUDE.md, I also manually execute /compact. "

claude.md Detailed records:

Configured MCPs (usage scenarios + invocation prerequisites)
Configured Skills (emphasize following the skill.md SOP process + required tool calls)

Why: AI system reliability does not come from the model, it comes from clear metadata. Models change default behavior every generation; if metadata constraints stay the same, behavior stays the same.

7. Cline: Using AI to write MCP (Posts 962, 971, 978, 2025-03)

Lao Huang self-built 2 MCP servers via Cline,all written by AI:

Obsidian MCP(read/create/edit Obsidian notes)
Replicate Flux MCP(call Replicate Flux to generate images)

Lao Huang's learning advice (Post 1084):

Developing an MCP Server is very simple → use Cline to let AI write it
If you want to learn code → go to github.com/modelcontextprotocol/servers find a Server you're interested in and read the source code

See newtype · MCP Protocol and Ecosystem Practice §2 MCPs built by Lao Huang for details.

8. Gemini 2.5 Pro codebase scanning magic tool (Post 1071, 2025-04)

Usage: Directly feed the entire project code to Gemini 2.5 Pro → ultra-large context(1M+) → chat to understand the entire project's running logic.

Typical scenario: Lao Huang used it yesterday to scan the Second Me project code and fully understood it.

Why:

Cursor uses RAG retrieval; Gemini directly fits the entire repository
Suitable for the understanding the project phase;implementing code phase still uses Cursor / Claude Code

9. Vibe Coding mental models

9.1 Vibe Coding definition (Lao Huang practice version)

"AI only takes over the work of writing code. Humans still need to take the leading role and make specific judgments."(Post 1451)

Key point: You don't have to write code, but the depth of your thinking cannot decrease.

9.2 Lao Huang's 1-month Vibe Coding experience (Post 1437)

"I've been working on this project for just over a month, from frontend to backend, and learned so much. AI doesn't just change the way of collaboration,the way of viewing and thinking about problems also changes.

Luckily, I'm not a programmer,so I don't carry that much baggage and bias, and can get started quickly,invading the expert domain as an outsider. "

9.3 Prophecy of Vibe Coding cleaning services (Post 1789)

"Saw a new business:Vibe Coding cleaning. When you've passed the product validation phase using Vibe Coding and start scaling for real, you'll very likely need this kind of service.Because code written by AI may contain many hidden pitfalls. It's fine when the scale is small. When the system gets more complex and features keep increasing, these pitfalls will surface."

Lao Huang stepped on them himself (Post 1789):

Prompt House user login verification and subpage flows suddenly became unstable (constant refreshes) after adding more subpages
Spent a lot of time troubleshooting with Codex
Lesson: No architectural preparation for SEO subpages + kept adding new things without periodic cleanup

9.4 Key anti-pattern (Post 1789)

"In Vibe Coding, you don't have to write code, but the requirements in other aspects are not lowered at all."

Meaning:

Architectural thinking still needs to be done
Periodic refactoring still needs to be done
SEO / performance / security / user experience still need to be done
Otherwise AI will accumulate more and more "barely runs" code → eventual explosion

10. AI Coding for "non-technical vs technical" people (highlighted Posts 933, 1010)

Lao Huang's honest judgment(for newbie product managers):

"Don't believe those UP hosts who say zero-foundation people can develop too. I can tell you very responsibly,you still need to understand programming and have programming thinking. Otherwise tools like Cursor will drive you crazy."

But at the same time:

"All my experiences these years have been JUST DO IT. Whether doing content or development,I absolutely never wait until I've learned everything before starting. Everything gets solved while moving forward. "

The two do not conflict:

You must have programming thinking(understand concepts like variables / functions / classes / async / API calls / databases)
You don't need to have programming experience(no need to memorize specific syntax)
Learning by doing is the path, but you must start from the most basic concepts(watch a few introductory Python videos before you start)

11. AI Coding directions that Old Huang is not optimistic about

11.1 General-purpose Coding Agent (Manus, etc.) (Post 941 featured)

Cannot survive under the crushing dominance of OpenAI / Anthropic / Google
"All model-based tinkering will be covered by upgrades in model capabilities"
Domestic media hypes Manus only because the "Chinese AI" narrative requires it

11.2 Running Cline on local large models (Post 593)

"The coding capability of local large models cannot match Claude 3.5 Sonnet, so I do not recommend spending twenty or thirty thousand yuan specifically for this. "

Conclusion: Subscribe to Cursor / Claude Code; even if API costs are high,it is still cheaper than hiring people to develop.

12. Lao Huang's current (2026-05) AI Coding stack

开发主战场:
  Claude Code(Skills + MCP + Slash commands + Sub-agents)
  Cursor(辅助 / Composer 编辑 / Plan Mode 规划)
  Codex(2026-04 起作为补充)
  GitHub Copilot(模型供应商角色)

UI 设计:
  v0.dev(产品早期)
  Keynote(后期,Go Fundamental 心法)

代码搜索 / 项目理解:
  Gemini 2.5 Pro(大上下文扫码)

代码评审 / 大型重构:
  Claude Code + Ultra think

API / 模型组合:
  Claude(规划逻辑)+ Gemini(大上下文)+ Sonnet/Flash(执行)

13. Anti-pattern summary

❌	✅
Using Cursor alone to do everything	v0 for design / Claude for architecture / Cursor for implementation, each handling its own segment
Stubbornly sticking to domestic AI Coding tools	"Stick to the best": Cursor + Claude Code + Codex
Running Cline on local large models to save money	Subscribe directly to Cursor / Claude Code; the efficiency is much higher
Letting Cursor improvise freely	Write requirements clearly and break them into small pieces,feed them in step by step
Retrying whenever Claude Code gives a wrong answer	Add `Ultra think` to trigger the 32K thinking budget
Vibe Coding without thinking about architecture	Architecture / performance / SEO thinking still needs to be done; you just don't write the code
Not updating CLAUDE.md	Update it after resolving each issue, and for major issues update it midway `/compact`
Waiting until "I finish learning programming" to start	You must have programming thinking; experience is accumulated while doing,JUST DO IT
Switching tools as soon as you see Augment / Trae	Consider that only after the current main battlefield is fully connected

References and extensions

Main references: All 315 posts in the ai-coding topic bucket. Covered post numbers (complete full set, covering 2024-08 to 2026-05): 471, 473, 477, 479, 480, 481, 482, 483, 486, 487, 488, 535, 545, 556, 577, 593, 630, 637, 639, 643, 646, 655, 713, 717, 722, 724, 731, 764, 794, 807, 818, 826, 838, 841, 843, 854, 862, 878, 882, 883, 893, 895, 926, 930, 932, 941, 962, 965, 971, 975, 978, 1000, 1037, 1071, 1079, 1083, 1084, 1150, 1154, etc.

Supporting wiki:

newtype · Vibe Coding Independent Product Practice — Prompt House / LearnFromYouTube and other actual products by Lao Huang
newtype · Agent Architecture and Multi-Agent Orchestration — Claude Code Sub-agents / Multi-Agent orchestration
newtype · MCP Protocol and Ecosystem Practice — Self-built MCP for Cline / Cursor + MCP combinations
newtype · AI Tool Stack Practice and Evolution — Holistic view of the tool stack

History

2026-05-16 v1.0 created — Phase 1.1 wiki #7. Based on all 315 posts in the ai-coding topic bucket (2024-08 to 2026-05), no sampling.

2026-06 supplement: Demonstration-style automation, research division of labor, and failure boundaries

The following is a June snapshot of tool usage in the newtype community, not the current official operation document. Prices, plans, model names, and entry points only represent the statements made in the posts at the time, and have not been re-verified online in this revision.

From verbal steps to demonstration, correction, and reruns

On the 19th–21st, the author introduced Record & Replay: record an operation once, and organize the steps into a reusable Skill. During his own demonstration, he missed one step; after it was pointed out, he updated the Skill and reran it before it worked correctly. A reusable method is not "record once and it will be forever correct", but rather select repetitive work → demonstrate → check for missing steps → revise rules → rerun with actual results. On the 21st, preparing to inventory repetitive work on the computer was still a plan.

He thus predicted that the RPA (automatically operating software according to pre-set rules) market will be replaced, and regarded Goals, Automations, Heartbeat, and Skills as an enterprise automation portfolio. The market growth rate and share in the post were not accompanied by verifiable full text; one reader simultaneously reported using a combination of RPA and AI and questioned reliability. Therefore, the product assumption that "maintenance methods may change" can be retained, but it cannot be written as "replacement is already complete" or "stability is definitely higher".

General Agents and research tools divide labor according to process

On the 22nd, the author's combination was Perplexity collecting materials and exporting Markdown, Codex then organizing, making suggestions, and calling content Skills, and newtype OS using the complete materials to help him draft structures and propose revisions; final choices and the finished manuscript were still completed by himself. Writing the framework and opening first, letting AI generate a version to see the overall effect, and then writing by himself uses the draft as a thinking sample, not automatic publication.

The June 3 Q&A and the 25th quota Q&A recorded contemporary statements such as Max having 10,000 monthly credits and the 35,000 bonus being one-time; the author additionally purchased about 3,000. He chose paid research tools to save himself the work of customizing financial data access for general Agents. The medical market research comparison on the 9th only stated that Perplexity's output was more detailed, and did not provide the two readable reports from this batch, so it cannot be concluded that its accuracy is higher. Both users who connected Health to Apple Health reported being unable to retrieve data; the issue was unresolved, and it cannot be inferred that upgrading the plan will fix it.

Observable results are more important than "the effect is good"

On the 10th, he exported Google/Bing data approximately every ten days for Codex to analyze; on the 14th, he suggested providing the keyword spreadsheet and classification Skill together; on the 27th, he reported that his Wiki GEO had taken effect. There are clear inputs and cycles here, but the accompanying images were not read, so growth amounts cannot be recovered nor causality determined. On the 8th, he used Goal to uniformly revise religious wording in a book manuscript; the author self-reported about two minutes and 83,363 tokens; this can only serve as a single task case, not a cost benchmark.

Verify compatibility and memory separately

The short post on the 17th about "can connect to open-source models" does not necessarily contradict the DeepSeek/Kimi configuration failure on the 19th: the latter involves API interface compatibility. The author said OpenRouter works, and comments suggested switching adapter tools; there are no reproducible configurations in this batch, so installation commands will not be fabricated. On the 21st he claimed Skills are migratable, but on the 22nd a member who tried them worried that long-term memory was weak; being able to copy files does not mean dependencies, permissions, and memory experience are all identical.

Only a preview of the long article about memory remains; what can be confirmed is that the author emphasized migratable memory across Agents and across projects. There is also a contemporary menu record of manually enabling memory on the 22nd, which is not guaranteed to be the current default setting. Help requests about account bans, repeated reconnections in long conversations, headless call costs, etc., have no complete resolution processes and are all retained as unresolved issues; it cannot be deduced from "switching tools might help" that this is a reliable troubleshooting solution.