Tag: Vibe Coding

I interact with LLMs daily, both at work and in my personal projects. For many of us in tech, leveraging these language models has become second nature.

It’s well-known that folks in the tech sector, like me, engage with LLMs at twice the rate of the general population. In my case, LLM usage often exceeds a full day each week.

Even as regular users, we sometimes find ourselves frustrated when an LLM doesn’t quite deliver the responses we expect. Here’s how I effectively communicate with LLMs during vibe coding sessions. These insights are just as valuable when navigating extended interactions with an LLM UI like ChatGPT.

Choosing My Vibe-Coding Environment

Vibe coding is the art of co-creating software with AI. I lay out my vision, the AI generates code, and together we refine it to match my intent. However, the process isn’t always smooth sailing.

The first step in my workflow involves choosing a coding environment. This space serves as a hub for interacting with the LLM, drafting, and executing code. I’m partial to Cursor, having started on their free Hobby plan, but I’ve since upgraded to the Pro+ account due to my extensive usage.

For those interested, here are some environment options:

• Cursor: Widely used by vibe coders for its customizable interface.
• Windsurf: An alternative that executes terminal commands independently.
• Google Antigravity: A unique option favoring agent-driven development.

In my examples, I’ll be using Cursor, but the principles are applicable across platforms. Even if you’re simply delving deep into LLM conversations, the same guidelines apply.

Why Prompting Alone Isn’t Enough

You might ask why we’d even need a tutorial for vibe coding. It’s true—the basic idea is simple: specify an outcome, and the LLM delivers. However, once the complexity increases, especially when dealing with multifile systems or tools, context management becomes crucial.

The context window is a pivotal concept. It’s the memory scope LLMs use to handle input/output data, a window defined by token limits. For example, GPT-5.2 allows a 400,000-token window, while Gemini 3 Pro goes up to 1 million. Understanding this helps in avoiding token overflow, which can diminish retrieval accuracy.

Expert commentator Matt Pocock explains the nuances of context windows well—view his YouTube video for more insight. For now, keep in mind that effective planning minimizes verbosity and assumes clear window management.

• One team, one dream. Divide projects into manageable phases, clearing LLM memory regularly between tasks.
• Do your own research. While you don’t need exhaustive detail, grasp general methods and potential build paths.
• Trust but verify during troubleshooting. Get clarifications from the LLM and cross-check details externally.

Explore Further: How Vibe Coding Transform Search Marketing Workflows

Tutorial: Creating an AI Overview Question Extraction System

To produce high-ranking content in AI Overviews, address the questions they respond to. This tutorial guides you in developing a tool to extract such questions, not just to provide a use case but also to demonstrate effective system development via vibe coding. It’s not a guaranteed path to AI prominence but offers strategic insights.

Step 1: Planning

Before diving into Cursor or any other tool, identify your goals and necessary resources. Although it’s early days, using generative AI for initial brainstorming can be beneficial. I often start by articulating my end goal in a sentence or two, alongside requisite steps, in AI tools like Gemini or ChatGPT. Missteps here are okay—this stage is about outlining thoughts, not finalizing builds.

For instance, I could outline:

I’m an SEO, aiming to leverage Google's AI Overviews to inspire our authors' content. We need to extract implicit questions addressed by AI Overviews. Proposed steps include:

1 – Choose a keyword target.
2 – Run a search and collect the AI Overview.
3 – Deploy an LLM to derive underlying questions from the AI Overview.
4 – Preserve questions in an accessible format.

With a clear direction, select your preferred LLM. While I’m partial to Gemini for chats, modern models with robust reasoning will suffice. Initiate a session, state your intent to build an AI Overview extractor, and share your planning prompt.

Step 2: Laying the Foundation

Cursor offers diverse models which I find advantageous. For this task, start in Plan mode, allowing for structured discussions and informed decision-making.

Kick off discussions with our defined project prompt.

Making modifications is crucial, so carefully review the LLM’s plan to ensure alignment with your vision. Address any disparities through collaborative discussions with the model.

Consider seeking insights into possible project failure points and implement preventive measures accordingly. For efficiency, I tend to request models to generate outline files for improved context window management, validating internal consistency before proceeding.

Step 3: The Build

With the foundation laid, shift to Agent mode using your selected model—in my case, Gemini 3 Pro—to execute the building phase. Keep an eye out for required approvals during script execution to ensure a smooth process.

Once script development is complete, proceed with library installations via the provided requirements.txt file. For organized dependency management, setting up a virtual environment is recommended.

Running your first script execution often surfaces unforeseen challenges. Tackle these by leveraging comprehensive diagnostic feedback, ensuring issues are resolved before moving forward.

Troubleshooting and Improvements

My initial run revealed a lack of expected AI Overview detection—a misstep rectified through close inspection of terminal outputs, model adjustments, and informed re-execution.

Embrace troubleshooting as a key growth component in the vibe coding journey, enhancing reliability and performance as you fine-tune system components.

Dive Deeper: Inspiring Examples of Responsible Vibe Coding for SEO

Logging and Output Management

Employ Weave for maintaining organized records of query inputs and LLM outputs. This robust tool aids in both immediate log assessment and long-term query-trace reference.

Use the analyze_query trace to monitor pivotal data points, fostering awareness of the direct connection between query intentions and AI Overview content insights.

Structure Over Vibes: A Strategic Approach

Across my years of vibe coding, I’ve learned structure creates reliability—increasing complexity demands methodical workflows, ensuring sustainable success. Remember to keep the vibes in your collaborations strong, united by a shared purpose and approach.

Inspired by this post on Search Engine Land.