Reported by New Intelligence Yuan
Editor: Hao Kun, Yuan Yu
【New Intelligence Yuan Highlights】 A top 0.01% elite player globally, once officially recognized by Cursor, decisively "defected" after the release of Claude Code 2.0! This isn't just a tool switch; the underlying rules of the game have quietly changed.
If you follow the AI programming circle, you've likely been flooded with discussions about Claude Code recently.
But amidst the noise, Silen Naihin's voice stands out for its thought-provoking depth.
He is no ordinary developer; he was a top 0.01% user of Cursor globally, once the most loyal advocate of its ecosystem.
However, after the release of Claude Code 2.0, he made a decision that shocked everyone: abandoning Cursor and fully migrating to Claude Code.
In a viral blog post, he documented this complete journey of his heart and mind.
This is a comprehensive practical guide from a "veteran" who has been exploring AI programming since 2021 and has read all the Claude Code guides. It includes:
- 5 years of accumulated experience in AI programming
- Practical insights from using Claude Code
- Compilation of over 10 essential articles and countless tweets about Claude Code
- Exclusive configurations
- Advanced techniques
After reading this, you'll truly get the hang of it; the only limit will be your imagination.
For the full detailed content, please see the author's blog original address: https://blog.silennai.com/claude-code
From GPT-4 to Cursor
A Journey of Disillusionment
The story starts in March 2023.
At that time, Github Copilot was at the forefront of AI programming, ChatGPT was still a novelty, and continuous model improvement was not taken for granted.
Then, GPT-4 was released, clearly marking a paradigm shift.
Silen Naihin recalls that this was the first time he began to fantasize about creating an "AI thinking loop"—allowing AI not only to write code but also to cooperate with tools and search the web to form a closed loop.
Later, people gave this loop a resounding name: AI Agents.
Silen Naihin was fortunate; he participated in building the first truly popular AI agent—AutoGPT.
This project quickly ignited globally and remains a legendary repository on GitHub that reached 100,000 stars the fastest.
But to be honest, it wasn't easy to use.
Silen Naihin admitted that AutoGPT at the time was not perfect.
At its luckiest, it could produce a barely functional Tic-Tac-Toe game; but if you wanted something more complex, it was simply a fantasy.
It wasn't until Cursor emerged in 2023.
Initially, Silen Naihin was also skeptical.
He tried it out in October 2023 and May 2024, and both times he gave up.
In his view, the old-fashioned "copy and paste from ChatGPT" was still more reliable than these half-baked tools.
Until September 2024, when Cursor Composer arrived.
At that moment, quantitative change led to qualitative change.
Silen Naihin found that 90% of his code was now handled by generative AI. He was like a man possessed, "living" in the editor, frantically exploring the limits of this tool.
He wrote an unpublished internal best practices guide, figuring out every tiny trick:
How to perform surgical cursor placement, how to finely manage the context window, how to write Cursor rules, how to probe the boundaries of the model's capabilities...
He thought he had found the "silver bullet". He even received an email from the Cursor team congratulating him on becoming a top 0.01% user.
And earlier this year, Silen Naihin actually tried Claude Code, but the result was—another abandonment.
At the time, he felt that Claude Code's workflow was a step backward.
The model wasn't smart enough, and most of the time, humans still needed to clearly understand what was happening in the code.
If that was the case, why use a command-line tool with barely adequate functionality and a user experience ten times worse?
The Comeback of Claude Code 2.0
Why Has the Abstraction Level Changed?
Claude Code 2.0 has arrived.
On the surface, its user experience has evolved, its shell framework is more flexible and robust, and many bugs have been fixed. But in Silen Naihin's view, these are just the tip of the iceberg.
The truth is, regardless of what Anthropic did to the underlying Opus 4.5 model with RLHF (Reinforcement Learning from Human Feedback), they have completely changed the rules of the game.
We have evolved to the next "abstraction level".
In the past, you needed to review code, instruct the model at the file and function level.
Now, Silen Naihin has discovered that you no longer need to stare at every line of code; you can directly test "behavior".
To prove this, Silen Naihin built a genetic algorithm simulator in just one day.
This simulator features an interactive visualization interface for real-time evolution, including complex functions like fitness functions, selection pressure, and mutation rate regulation.
The key point is: he didn't write a single line of code by hand.
Not just him, others are also using this "magic wand" to create miracles:
Someone made a visualization plugin for office work, someone made a complex bionic creature simulator, and someone even made a $30 smart bird feeder camera.
Silen Naihin himself couldn't stop, building the Wright Brothers flight simulator, a personal website, and the interactive blog article you're reading now.
The magic wand is now in everyone's hands; the question now is: how will you wield it?
On Twitter, a skeptic asked: "Can someone explain why use Claude Code instead of Cursor?"
Until a month ago, Silen Naihin had the same question. But now, he has a definitive answer:
First, Asynchronous First.
Staying in the IDE, humans instinctively go for code review, getting caught up in perfect details, but this is low-dimensional diligence.
Claude Code elevates the workflow to the next abstraction level, and the terminal-native workflow is the key that forces you to take this step.
Second, RLHF for Its Own Scaffolding.
The Claude model (especially Opus 4.5+) performs significantly better in the Claude Code environment.
File search, tool usage—everything has been optimized for this interface. It's like playing on its home turf, feeling right at home.
Third, Cost Efficiency and Customizability.
It has a higher cost-performance ratio, and DIY is native, with composability built-in. It's not a closed black box, but an open Lego world.
Of course, Silen Naihin also objectively pointed out that Cursor still has its place.
If you pursue pixel-perfect frontend details, or if you program for the sake of learning programming, Cursor's tight feedback loop is still optimal.
But if you only care about output, and if you, like Silen Naihin, are an "abstraction maximilist," then Claude Code is your ultimate destination.
Ultimate Configuration and the Secret to "One-Click Mastery"
Before diving into specific tactics, Silen Naihin also generously showcased his arsenal:
- Claude Code with Opus 4.5: Responsible for planning, code generation, and architectural decisions—the "heavy lifting";
- Cursor with GPT 5.2/Sonnet 4.5: Used for fine-tuning UI and small-scale changes;
- ChatGPT: As a "second opinion" consultant, providing different perspectives;
- Ghostty: As an ultra-fast terminal;
- Wisp: A voice input神器, saying goodbye to "mouse hand."
To make it easy for everyone to get started, Silen Naihin did something geeky:
He encoded the core secrets (Alpha) of this article into two commands:
/setup-claude-code (global) and /setup-repo (project-level).
They will understand your needs like an interviewer and automatically configure everything.
Download these commands to ~/.claude/commands/:
- /setup-claude-code (run once per machine): https://gist.github.com/SilenNaihin/3f6b6ccdc1f24911cfdec78dbc334547
- /setup-repo (run once per project): https://gist.github.com/SilenNaihin/e402188c89aab94de61df3da1c10d6ca
Agent Programming in Practice
The Five Pillars and Efficient Techniques
To truly master Claude Code, Silen Naihin summarized the five pillars you must master.
1. Context Management: Don't Let It "Forget"
Although Claude Code is powerful, it has a 200k context limit, making it more prone to "hitting a wall" than other tools.
Silen Naihin reminds: Staying focused is key; a chat (session) should ideally correspond to only one task.
Learn to use the /compact command. Although there will be some loss, it's usually worth it to stay in the same chat context. If the context is truly insufficient, use /transfer-context to gracefully transfer the context.
/transfer-context: https://gist.github.com/SilenNaihin/e4be0e8750343d9cbafdaab88366115c
Silen Naihin also has an exclusive insight: Generating documentation or tests in an existing chat context always yields the best results.
2. Planning: Sharpening the Axe Doesn't Delay the Work
Spending 1 minute on planning can save you 3 minutes later.
For large tasks, press Shift+Tab twice to enter planning mode, or directly formulate a plan through conversation in the chat.
After creating a plan, use the /interview-me-planmd command. Let Claude act as a strict product manager and conduct an in-depth interview with you to ensure nothing is missed.
Silen Naihin specifically mentioned that Opus 4.5 is particularly good at explaining logic and drawing ASCII charts.
/interview-me-planmd: https://gist.github.com/SilenNaihin/0733adf5e8deea4242878938c3bdc9fb
Also, remember to clearly tell it: Don't be timid just to maintain backward compatibility (the model is sometimes too "obedient" now), and beware of over-engineering—we only need the simplest changes, and code readability is king.
3. Closed Loop: The Art of Automation
The old programmer joke about "spending a week to automate a 5-minute task" is outdated. In Silen Naihin's view, completing a closed loop is almost always worth it now.
If you find yourself repeating something, automate it through commands, agents, or updating configuration files.
Don't even let trivial things like modifying tsconfig slip by.
4. Verifiability: Trust, but Verify
You no longer need to review code line by line, but you need to verify behavior. This means interface testing, looking at the UI, clicking buttons, sending API requests.
For large-scale refactoring, Silen Naihin suggests letting Claude pre-build comprehensive interface tests—this is your safety net. As long as the integration tests pass, you can boldly release.
In Silen Naihin's words: "Let Jesus take the wheel!"
5. Debugging: Be a Cold-Blooded Detective
How to fix bugs in AI-written code? Don't panic; debug systematically.
Silen Naihin developed a /debug command specifically for generating hypotheses, reading code, and adding logs.
/debug: https://gist.github.com/SilenNaihin/6833c01f597c82912af5aca4e3467a35
Follow the "three strikes" rule: If it still doesn't understand after three explanations, change your strategy and show it examples instead of just talking.
If it really doesn't work, form a "model committee".
Use /ensemble-opinion to let Claude, Gemini, and Codex diagnose in parallel, or directly use /codex-delegate to let Codex take over.
/ensemble-opinion: https://gist.github.com/SilenNaihin/3e9b43522b61e155bd256fe7193493cd
/codex-delegate: https://gist.github.com/SilenNaihin/ff19b2d65d17137b0ee1f609f25205c5
In addition to these five pillars, there are some small tips that will double your efficiency:
- Keyboard Shortcuts: Esc+Esc is the regret medicine (rollback), Ctrl+R is the time machine (search history), allowing you to run bash commands directly in the chat.
- The "Just Ask Claude" Mindset: Don't manually change configurations yourself; just command it to do it. It will search for what it doesn't know.
- Multithreaded Operation: Open 12 terminals at once. Don't be afraid of the "blast radius"; hitting the same branch hard is usually faster and simpler than using git worktrees. As long as you assess the risk, the worst-case scenario is just a rollback.
Domain Playbooks and Advanced Plays
From Frontend to AI Research
For different domains, Silen Naihin offers different strategies:
- Frontend: Good prompts and guidelines are key. Screenshots are神器; drag them into the chat, and Claude can understand them instantly. Remember to install the Vercel React best practices skill and make the linters shut up (or fix them).
- Backend: Use ORM as architecture-as-context, invest in realistic seed data, and let AI self-verify.
- AI Research: Give Claude access to a VM with A100s; it can run experiments, read logs, and do analysis itself. As Andrej Karpathy said, watching it coordinate work beside you is an absolutely new experience.
Advanced players can also try Ralph (for large projects, but to be honest, it's a bit troublesome), mobile programming (via vibetunnel.sh), and various Hooks, Subagents, and MCP.
Especially the headless mode (-p), which allows Claude to silently perform code review or respond to tickets for you in the background.
Ralph
Finally, CLAUDE.md is the soul of your project.
Generate it with /setup-repo and keep updating it based on pain points. Give Claude clear directory structures, dependency patterns, and non-standard choices, and it will reward you with more precise code.
/setup-repo: https://gist.github.com/SilenNaihin/e402188c89aab94de61df3da1c10d6ca
Everything will change, and it will change quickly.
But Silen Naihin believes that some truths are eternal:
The leverage of planning will only increase, the ability to verify quickly remains important, completing a closed loop is the core of automation, and most importantly—don't be lazy.
Figure out what needs to be done logically; those brain-burning thoughts are your irreplaceable value.
There is no so-called "standard answer," only the system that suits you best, found through continuous experimentation and iteration.
Now, take this guide and go create your own miracles!
References:
https://blog.silennai.com/claude-code
Instantly Chase ASI
⭐Like, Repost, and View - One-Click Three Links⭐
Light up the star, and lock in New Intelligence Yuan's fastest push!