James Dooley: Vibe coding. The amount of people that reach out to me asking what vibe coding is. Today I am joined with Andrea Abbondanza, who has been doing a lot of vibe coding recently. He is a specialist in SEO for restaurants. So for anyone who is watching this, Andrea, what is vibe coding to start with?
Andrea Abbondanza: It is coding with vibes. That is what it is. I feel like I am a full stack senior developer. I am using Claude Code and recently I have been implementing and using Claude Code inside Anti Gravity. I mainly use Claude Code for creating the whole process, building agents and orchestrations. Then I use Google, especially Flash, for debugging. I have not slept properly for the last three days, James Dooley. I got so excited. It is great. With vibe coding, I have been able to create projects that I had been waiting weeks or months to build because I can now do it myself.
Andrea Abbondanza: The tool helps you build things, but it does not just act like an LLM telling you what to do so you copy and paste. It pushes things live and then debugs them. Debugging used to be the main bottleneck for developers. Now that problem is largely removed. You can integrate GitHub and multiple technologies. It is crazy. I was able to create a full online AI avatar yesterday. The hardest part was integrating Facebook and YouTube APIs because that takes time. But vibe coding has been powerful. It feels like learning something new at speed again. The excitement is unreal.
James Dooley: So what is it that you are actually building then? Are you vibe coding apps, websites, programmatic SEO systems, or something else?
Andrea Abbondanza: The first thing I am building is AI avatars. You have probably seen them gaining traction. It is usually a persona. You can use tools for voice like ElevenLabs, image generation tools, and then connect everything together. The whole workflow can be created with vibe coding. That is one project I am working on.
Andrea Abbondanza: The other part is memory. LLMs hallucinate when they handle too much data because their memory is limited. So you need to create agents that handle different processes. For example, one main agent acts as a marketing manager that handles operations. Then you have a social media agent that handles content and publishing. Another agent handles distribution on platforms like YouTube and Instagram. Then you connect APIs to make everything work.
Andrea Abbondanza: Each agent has a defined role. The orchestrator is another agent that manages all of them. It acts like a supervisor. When each agent is specialised, you get a full workflow that actually works. That is the system I am building. My friend Arie is an engineer and we have been working together. He replicated an app yesterday, built a website, and it can be sold at a fraction of the cost. That is the shift. Things that used to take teams can now be built in days.
James Dooley: That is a big shift. So when it comes to LLMs for vibe coding, what do you prefer? Is it GPT or something else?
Andrea Abbondanza: I mainly use Claude Code. I am running a workshop soon and I want people to build things live. Claude Code allows you to push changes instantly. I also test Anti Gravity, which is a Google platform that works with different LLMs. You can use OpenAI, Claude, and Gemini. Gemini Flash is very fast but runs out of context quickly.
Andrea Abbondanza: The best setup is combining tools. Claude Code builds the structure. Gemini helps with debugging. You can use multiple LLMs at the same time. That is where things get powerful. I have not used OpenAI as much for this because of cost and limitations, but it still has its place.
James Dooley: We noticed when testing that context windows matter. Claude handled context better and reduced hallucination. Are you using Perplexity as part of your workflow?
Andrea Abbondanza: I use Perplexity more for research. It is good for real time data. For example, if an agent needs current information, you use Perplexity to fetch it and feed it into the system. What works best is assigning different LLMs to different tasks. If GPT writes better content, use it for that. Then use another model for review. Then have another agent enforce rules and output quality.
Andrea Abbondanza: When you get a result you like, ask the LLM to generate the prompt that created it. Then you can reuse that prompt. That reduces hallucination and improves consistency.
James Dooley: That is smart. You are using AI to improve prompting, then building agents that handle prompting automatically. It removes a lot of manual effort.
James Dooley: Anyone watching this, are you vibe coding? What are you building? Leave a comment and share what you are working on. If you want to come on and talk about your setup, your agents, or your preferred LLMs, reach out. Things are moving quickly and it is hard to keep up. Andrea has not slept for three days because he has been deep in this.
James Dooley: Andrea Abbondanza, it has been a pleasure. Thank you.
Transcript source: Provided by creator in RSS feed: download file