Software Engineering :: Pair Programming :: AI Assistant :: RepoPrompt :: About :: Ultimate Vibe Coding Setup: o3 pro, Cursor, Codex, Claude Code
⪼ Made with 💜 by Polyglot.
- AI :: Dictation :: Software :: superwhisper (
⤴︎) - AI :: OpenAI :: Codex (
⤴︎) - Software Engineering :: Pair Programming :: AI Assistant :: RepoPrompt :: About :: It's all about what context you provide into the models (
⤴︎)
This video is a tutorial + workflow demo designed to educate and inspire developers on how to become highly productive using a “vibe coding” stack. The creator showcases his personal workflow using a set of advanced AI coding tools (Cursor, Codex, Claude Code, O3 Pro, SuperWhisper, RepoPrompt, and more) to build his AI startup, Vectul. The tutorial walks through real coding tasks, illustrating how he uses AI agents collaboratively and strategically for code generation, review, and task automation. The key takeaway is that mastering both the tools and the learning mindset is crucial for future-proofing your developer skill set.
- Cursor: The primary AI-powered IDE used for vibe coding.
- Claude Code: Used for deep code reviews and inline assistant support in Cursor.
- Codex: OpenAI's asynchronous coding agent that excels at multi-version task handling.
- O3 Pro: Great for complex bug resolution and high-risk refactoring with context.
- RepoPrompt: Helps build optimal prompts by attaching the right files and instructions.
- SuperWhisper: Transcribes voice into prompt-ready instructions for seamless flow.
- Uses Codex to solve tasks with multiple versions, selects best one, then submits PR.
- Claude Code reviews PRs with senior dev mindset, sometimes using two agents simultaneously.
- RepoPrompt and Claude help generate context-rich prompts for O3 Pro to assess advanced tasks.
- Cursor’s agent and Codex used interchangeably depending on use case and latency.
- Leveraged Gemini 2.5 for summarization and additional model diversity in decision-making.
- Always use a dedicated branch for AI-generated PRs, never merge directly into production.
- Use “do not change anything else” prompts in agent mode to maintain tight control.
- Use “answer in short” in ask/plan mode to learn and clarify from AI assistants.
- Maintain annotated markdown prompts and
.mdcconfig files to guide agents effectively. - Structure long prompts with XML-style tags like
<user_instructions>for better parsing.
- Codex has the highest pull request approval rate among cloud agents (85%).
- Cursor’s background agent is second, but still lags behind Codex in success rate.
- O3 Pro is powerful but requires more setup; not yet fully integrated with rest of stack.
- “Vibe learning” is more important than vibe coding—don’t just automate, understand.
- Develop taste and discernment by cross-validating tools and manually reviewing outputs.
- If your AI tool usage doesn’t scare you financially, you're probably not pushing hard enough.
- Productivity is multiplied by using agents in parallel, not in isolation.
- Final recommendation: If stripped to just two tools, he’d keep Claude Code and Cursor.
- Built Vectul (an AI-native productivity tool) from zero to $10k+ MRR using this workflow.
- Promotes joining The New Society, his paid membership offering templates, prompt libraries, and deep-dive courses on these tools.
- Encourages developers to master these systems now, before they're mainstream.