Notes on Vibe Coding

Introduction to Vibe Coding

It's a new way of coding with the help of AI tools. After using lots of tools, I am currently happy with Claude Code, which gives me better results, without going back and forth to my browser or code editor.

But then there are multiple tools, like OpenAi Codex, and the most famous one like Cursor (with built in editor).

Vibe Coding, Context Engineering, Spec-Driven Development

As a consumer or a normal user, I would like Vibe Coding to define like - you just have an idea, you just want to get quick output without even getting into details of how you want to be make it.

Context Engineering comes into play when you are not getting the output you are expecting AI to give you. And you want to control aspects of it, giving specific instructions or mostly accurate inputs to work on.

Spec-Driven development is a way of working with AI, just like old times, where you give a detailed spec of what you want to build, and AI will help you build it. The action mode comes at very end, and most of the Human time goes into planning and defining what needs to be done.

Claude Code Masterclass

Create Your Own AI Factory (Vibe Coding at Scale)

Source: Mark Kashef - AI consultant and educator specializing in prompt engineering

Most people use Claude Code for small tasks, but it can be used at a much larger scale. Here's a four-act blueprint to create your own "AI Factory":

9 Multiplier Hacks to 10x Output

Use #memory tag - Update the project's command center (Claude.md) to store persistent context
Leverage plugins - Bundle instructions and MCP servers for efficiency
Manual MCP config - Configure via .mcp.json for better reliability than auto-setup
Agents for tasks, not roles - Use AI agents for specific tasks rather than general roles
Use /rewind command - Version control for your Claude sessions
Implement sanity checks - Use plan mode and tracking files before executing
Use /compact and /clear - Manage context window to prevent "AI slop"
Ask clarifying questions - Instruct Claude to ask before destructive actions
Break into phases - Use tracking files (e.g., ONBOARDING_TRACKER.md) for complex projects

Multi-Model Mastery

Use Claude Code within Cursor to access different models on the same codebase
Use web-based Claude for on-the-go tasks
Combine strengths: Claude for architecture, faster models for implementation

The "Replicate Anything" Framework

Build complex tools by:

Feed Claude raw API documentation as markdown
Learn from mistakes and iterate
Examples: branded PowerPoint generator, video analyzer

"Claude Code is not a magic wand; things will break as you're building. Just be comfortable with that."

"In the age of AI, anyone can do anything. If you're ever lost or confused, you always have someone there to ask, and you learn by doing."

How Claude's Brain Works (4 Layers)

System Prompt - The foundational instructions
User Memories - What Claude remembers about you
Conversation History - The ongoing context
Your Current Message - What you're asking right now

Every message you send gets processed through all 4 layers. Understanding this helps you communicate better with Claude.

Claude Code with Opus-4.5 Features

Key capabilities to learn:

MCPs (Model Context Protocol) - Extend capabilities with external tools (11:04 in tutorial)
Skills - Reusable prompts and workflows (28:00)
APIs - Integrate with external services (38:00)
Documentation - Auto-generate docs from code (44:12)
GitHub Integration - PR reviews, commits, branch management (1:08:08)

Best Prompting Techniques

10 internal techniques used by OpenAI, Anthropic, and Google:

Chain of Thought - Ask the model to think step by step
Few-Shot Learning - Provide examples of desired output
Role Prompting - Assign a specific role/persona
Structured Output - Request specific formats (JSON, markdown)
Decomposition - Break complex tasks into subtasks
Self-Consistency - Generate multiple responses and pick best
ReAct Pattern - Reasoning + Acting in loops
Tree of Thoughts - Explore multiple reasoning paths
Meta-Prompting - Ask the model to improve its own prompts
Constitutional AI - Add ethical constraints

Practical AI Coding Workflows

Building with AI Agents - The Right Way

Step 1: Use CLI over web (it's more capable) Step 2: Feed context, then switch to spec mode Step 3: Question everything like a philosopher Step 4: Link docs + repos for the agent to explore Step 5: Let Opus rip with autonomy

The "Ralph Wiggum" Method - Ship Features While You Sleep

Write a detailed PRD (spend an HOUR on this - it's worth it)
Convert it to small, atomic user stories
Add clear acceptance criteria for each story
Loop your AI agent through each story
It logs progress and handles edge cases

Scale Solo Business with AI Agents

Use AI agents for each business function:

Engineering: code, testing, DevOps
Design: UI/UX, brand assets
Marketing: content, SEO, social
Sales: leads, outreach, demos
Support: tickets, docs
Data: metrics & insights

THE KEY: You manage LLMs. LLMs manage tasks.

RAG 2.0 Approach

Building "RAG 2.0" is essentially making Claude Code run over your filesystem. To make this work:

Virtualize your filesystem to prevent the agent from messing stuff up
Use AgentFS by Turso as an example
Give the agent controlled access

Tech Stacks That Work

$130K Revenue Stack (T_Zahil)

Frontend: Nuxt, Nuxt UI, Tailwind CSS
Backend: Nuxt.js
Auth + Database: Supabase
Payments: Polar
Mails + Marketing: Bento
IDE + AI: Cursor Pro

Calculator App Making $300k/month - How to Clone It

Build the app with Rork
Use Pinterest for UI reference
Have ChatGPT turn it into a clean Rork prompt
Polish with follow-up prompts
Add monetization

Typeform Clone in Under 1 Hour

Using Claude Code Opus 4.5, built:

Full Typeform functionality
Spotify's design system
Working backend
User login (BetterAuth)
Database integration

AI Video & Content Creation

Making AI Videos That Don't Sound Like AI

Problem: Sora 2, Veo 3.1, Kling 2.6 all have that robotic AI accent.

When people swap in ElevenLabs, it sounds like:

A $5,000 mic
A soundproof studio (too polished = fake)

Solution: Match voice quality to content style, use natural-sounding techniques.

AI Influencer System (3 Steps)

Create an avatar with Nano Banana
Generate the script with ChatGPT
Turn script + avatar into UGC with Arcads

Cost: Fraction of paying $1,000 per video to traditional creators.

Contact Sheet Prompting in Nano Banana Pro

For fashion/product shoots - create consistent character across multiple frames using contact sheet format.

SEO Workflow with AI

The most underrated SEO workflow:

Start in Perplexity - Get complete SEO breakdown in 30 seconds
Send to Notebook LM - Turns into clusters, outlines, angles, expert insights
Push to Gemini - Get full blog, titles, meta description, TikTok scripts

Google Colab + VS Code Integration

Google now lets you run Colab notebooks directly inside VS Code, Cursor, Windsurf:

Free GPUs
Full IDE features
Better debugging
No browser restrictions

Tools & Resources

Learning

Anthropic's free 2-hour masterclass on Claude Code
Prompt Engineering Interactive Tutorial (GitHub)
50+ Claude Code skills collection (GitHub)
Claude Code guide by @bcherny

AI Agents

Do Anything Agents - Work independently for weeks/months, have their own email, self-manage projects
DeepTutor - AI learning assistant (1.4K+ GitHub stars)
Scouts - Monitors entire internet 24/7, messages you when something matters
Conductor - Fixes Claude Code review process friction

Development Tools

Visual workflow editor for Claude Code automation - cc-wf-studio
Grably - Transcribes and downloads media using Whisper AI
Pluely - AI assistant for meetings and screen shares