Get live statistics and analysis of CJ Zafir's profile on X / Twitter

vibe coding:

An Open Source AI model {Deepseek r1} launched 72 hours ago. This Chinese AI model just broke the internet. People are running r1 model on: - self hosted servers - mobile phones - raspberry pi - mac books Here're all the cool experiments, you check: 🧵

I code with AI tools for 6-7 hours daily. Built over 36 projects in last 12 months. Truth: "Build me ........... app" in 1 prompt is not possible. So, here're all the MISTAKES you might be making with AI code:

o3 is a research model Sonnet 4 is an agentic model Gemini 2.5 pro is a coding model Gemini 2.0 flash lite is a RAG model o4- mini is a task planning model Deepseek r1 is a thinker model GPT-4o is a chat model These are all the AI models you need to care about. Rest is fluff.

Ok here's the new flow: UI → Gemini 3.0 pro Code → Claude Sonnet 4.5 Plan → GPT 5.1 Just try this; remove all others.

never thought about it like that,

I've built 19 projects with Cursor AI without line a single line of code myself. But, the Truth is Cursor is dumb you don't add detailed docs around your project. You need to build a strong <Context Boundary> around Cursor Here what you can do to improve your Cursor workflow🧵

Cursor Prompting HandBook: 🧵

How to make Cursor Agent super powerful? A complete guide: ↓

I built this UI design with v0 and Claude·ai (web app) It took me 25 minutes to code this interface. Let me share my new UI design workflow: ↓

Massive Opportunity: If you know how to "fine-tune" AI models - it's your time to make $. Just fine tune Deepseek r1 for companies! This model is the first open-source model that is consistently good on local servers. Charge $20k to $50k for each fine-tuning job.

Here's the Master Trick to work with Cursor and Windsurf. It will remove 80% of AI hallucinations and 60% less code errors. You need to create a 'Documentation" folder in the root directory. Then you need to attach all your project docs to that folder Let me explain....

If you're using Cursor, Windsurf, Cline, Bolt, Lovable, Replit or v0. And you don't know what these docs are: - PRD - app-flow-doc - tech-stack-doc - frontend-guidelines - backend-structure - security-checklist etc You're not gonna make it.

Cursor/Windsurf technique to remove 80% hallucinations. Stop AI to think and plan. Attach 'coding documents' and force AI to use it as a knowledge base. If you code with AI, read this: ↓

vibe coding:

Ok here's the new flow: UI → Gemini 3.0 pro Code → Claude Sonnet 4.5 Plan → GPT 5.1 Just try this; remove all others.

o3 is a research model Sonnet 4 is an agentic model Gemini 2.5 pro is a coding model Gemini 2.0 flash lite is a RAG model o4- mini is a task planning model Deepseek r1 is a thinker model GPT-4o is a chat model These are all the AI models you need to care about. Rest is fluff.

I am convinced. I can start a new X account and take it to 10k followers in 60 days. You want me to share the insights with you to copy?

Some secrets: 1. Opus 4.5 > Opus 4.6 2. Codex 5.4 1M context is actually 512k 3. Qwen 3.5 is best OS model series (RN) 4. MCPs are dead 5. Openclaw was a prototype 6. Apple is the new NVIDIA 7. Perplexity is biggest grifter 8. Paper. design > figma 9. Ollama is just a crappy wrapper 10. Open source AI is ahead already.

After spending 500+ hours on ChatGPT, I can share 1 Ultimate Prompt Engineering Secret with you. Yes, for Free. You don't need any more prompt templates after this one! So here it is! 1st Step: Give it a role. "You are Spartacus, a pro copywriter," 2nd Step: Define its job "Your job is to write copy for online media brands" Third step: Give ChatGPT the control: "Now ask me all the questions that you need to know to write a perfect copy that'd align with my brand." **What you did here is rather than you figuring out the perfect prompt and context, you asked ChatGPT to do the hard work. Now it will give you 8-10 Questions, answer them in detail. By doing so, you are giving ChatGPT some data to be trained on. That's the trick. Now, to make it even better do this. 4th step: Define input: "Users will ask you to write copy for their business, brand, or anything else. It's your job to ask questions first to clarify the task and then implement the copywriting process. " Step 5: Define output: "You need to write in ______ style, start with the headline as a hook, write only ______ words per copy, make the copy persuasive, and use simple language" Step 6: Share 3-5 examples of inputs and outputs. That's it. Your ultimate prompt is ready. Here's how it looks. 1. Role 2. Job 3. Paste all question answers as context 4. Define input 5. Define output 6. Share 3-5 examples Now you'll get 50x better results. 80% of AI tools are built on this prompting technique. Thank me later.

I code with AI tools for 6-7 hours daily. Built over 36 projects in last 12 months. Truth: "Build me ........... app" in 1 prompt is not possible. So, here're all the MISTAKES you might be making with AI code:

I code 7-8 hours with AI every day. 6 days of the week. I tested all prominent tools in AI tech. - v0 - Cursor - Bolt .new - Replit Agent - Gptengineer - Windsurf IDE Here's my latest Ranking: Most improved project in 14 days: Bolt. new - 8/10 I am impressed by the output quality of bolt. Just 2 weeks back it felt half baked. Now it is shaping into something good. Great for coding Landing pages and micro saas (4-5 pages) Most Versatile Project: Replit Agent - 7/10 After Replit introduced React + Vite. It is now building applications with cool UI. Still a bit slow. A bit over complicated BUT moving in right direction. I coded a modern landing page in 1 shot that i designed in v0. Best in their domain: Cursor - For coding V0 - For frontend design Half baked projects: Windsurf IDE - 4/10 Disappointed with the output. It was breaking after 5th step every time. Still need improvement: GPTengineer - 4/10 Cool idea but competition is now super tough. Bolt is moving at crazy pace. v0 improved context with context caching. GPTengineer is still implementing 1 edit in one message. Here's my tech stack as of today. I use ChatGPT search for research I use o1 model for documentation I use v0 for frontend design I use claude AI for coding brain I use cursor for code implementation I use replit to deploy projects I feel like I can build any web app now with this stack in 5-7 days. I'll be sharing some cool experiments in coming days. Stay tuned.

An Open Source AI model {Deepseek r1} launched 72 hours ago. This Chinese AI model just broke the internet. People are running r1 model on: - self hosted servers - mobile phones - raspberry pi - mac books Here're all the cool experiments, you check: 🧵

98.77777% people in AI don't even know what is RAG. You're still early!

never thought about it like that,

Guys, what are you building? Let me help you vibe market that! Share your startups below and ask me anything.

Here's my AI tools cost right now: ✅️ Claude AI → $20 ✅️ ChatGPT → $20 ✅️ Bolt .new → $100 ✅️ Cursor AI → $20 ✅️ Gpt4o API → $110 ✅️ Claude API → $135 ✅️ Replit → $20 Unsubscribed: ❌️ v0 → $20 ❌️ Perplexity → $20 Looking into: 🟡 Lovable 🟡 Windsurf I use Claude × Cursor most of the time. But here's a breakdown: - For frontend components, I was using v0, and it was perfect (before v0 became full stack) Now it's not even 20% efficient. So, I had to unsubscribe. Now I am using bolt for that purpose. Bolt is super expensive. I think they need to cut cost by atleast 50% otherwise devs won't use it for production. Lovable is emerging as a great v0 replacement. I tested it few times and got good results. - For IDE, I am using Cursor AI from the start, but I see a cult community of Windsurf. I wonder why? Maybe the product has improved from when I tested 3 weeks back. So I'll check Windsurf and compare it with Cursor. Cursor Agent is amazing, but it's like a double-edged sword. If you don't have crystal clear instructions, it'll overkill your project. Proper Documentation + structured .cursorrules file is a must with Cursor Agent. - I use Claude and GPT-4o API in my products that this cost is from API usage. It will increase with time. - After the launch of ChatGPT search, I have a widget on my homescreen. I get same results as what I was getting from perplexity. And Perplexity gives you 3 pro searches per day. Which is enough if you want to compare anything. I don't remember when I used Google search last time. - I am not using Replit Agent much. I think Replit is the most overcooked AI product. Unnecessarily slow and overwhelming. I just use it to connect to cursor and test deployment quickly. Is there any tool that I am missing? That can help me code fast. I am coding 1 web app/a week. Production ready web app. ✅️ Fully responsive frontend ✅️ Backend integrated properly ✅️ Auth done ✅️ Testing done ✅️ Landing page When I look back, I get amazed at how I was building products just 3 months back. 2 months in development, super slow process, lazy devs, no control. Now, with Army of AI Agents and lean team. Tables are turned. And it is about to get crazy in few months. AI agents will be coding end to end products, and smart brains will be selling/renting these AI agents to companies. Anyways, that's my tool stack for now. Let me know what tools you are using in your workflow. Peace CJ

Massive Opportunity: If you know how to "fine-tune" AI models - it's your time to make $. Just fine tune Deepseek r1 for companies! This model is the first open-source model that is consistently good on local servers. Charge $20k to $50k for each fine-tuning job.