Opinions expressed by Entrepreneur contributors are their own.
Key Takeaways
- Vibe coding transcends small projects, with tech giants competing to master multi-file AI-augmented programming tools.
- Structured maintenance and smart input crafting result in significant time and cost savings in complex machine learning projects.
We tend to think vibe coding is only for people who can’t code, or that it’s limited to small projects. But it’s hard to talk about hindsight when the story only began earlier this year. Right now, OpenAI, Anthropic, Meta and Google are locked in fierce competition, investing billions of dollars to launch revolutionary AI tools that are all designed to handle multi-file projects.
As an engineer applying AI tools to complex projects, especially in machine learning, I’ve managed to cut both time and costs with three key solutions.
GitHub as the repository
The temptation with vibe coding is to accept whatever the AI generates and move on. That can lead to “zombie code” and untraceable bugs. The safest practice is to keep all source files and tests in a managed repository. In that case, I use GitHub to automatically run my tests on each commit and highlight indentation errors and broken functions. If a test fails, GitHub’s AI assistant points me to the offending line and offers a suggested fix. In my experience, this saves hours of searching through misaligned tabs.
To make this work, you should set up continuous integration from day one. Choose a default branch (usually main), configure automated tests and enforce pull‑request rules in GitHub. When a test fails, ask the AI assistant to explain the conflict before accepting any changes. This approach keeps AI in its proper role — as an assistant, not an overlord — and ensures your codebase stays clean and maintainable.
OpenAI’s agent mode as the crafter
You probably use ChatGPT at least once a week. And most of the time, you don’t get the perfect response on the first try. Why? GIGO — Garbage In, Garbage Out. The quality of your output depends entirely on your input. According to Google’s The Keyword, “before using a vibe coding tool, start with Gemini and try writing prompts that describe your ideas.” Large language models (LLMs) are only as good as the instructions they receive.
OpenAI is among the best for crafting effective prompts. I log in ChatGPT and use agent mode, which connects to both GitHub and real-time web search. With this setup, I can ask ChatGPT for a context-aware technical prompt by writing a trigger instruction such as: “You are a fitness coach tasked with finding the top five chest exercises in 2025. Research them on the web, then generate a technical prompt that Claude Code can use to create a landing page about them. Also, include my existing leg exercises from the workouts file.”
Anthropic’s Claude Code as the coder
Once you’ve crafted a technical prompt, hand it off to Claude Code, using Opus 4.1 for precise coding in multi-file projects and Sonnet 4.5 for quick technical questions. These models can generate Python, JavaScript and other languages with remarkable precision.
However, be mindful of Claude’s limitations: While it connects to GitHub, you must manually select and upload files from your repository. Once you reach the storage limit, you’ll need to start a new conversation and re-add your context. When that happens, don’t cut corners because the prompt is king.
OpenAI’s Codex as the inspector
OpenAI also offers a cloud-based engineering agent, Codex, that connects to GitHub in real time. With Codex, I can import any file from my repository simply by naming it. The tool automatically detects issues, visualizes them clearly and even provides a “Create extraction request” option to GitHub if I prefer not to fix the problem myself. In that case, Codex generates a dedicated branch and requests my approval before merging it into the main branch.
Make sure your main branch is set as the default and that your GitHub rules are properly configured to control merges and maintain code quality.
Here’s how my AI workflow typically unfolds for a multi-file project, step by step:
- Create the prompt: I start by asking ChatGPT (in Agent Mode connected to GitHub and real-time web search) to generate a context-aware technical prompt.
- Generate the code: I copy and paste the prompt into Claude Code using Opus 4.1 for precise, large-scale coding.
- Test and iterate: I copy or download the Python file, upload it to GitHub and commit to trigger automated tests. If a test fails, I resolve the conflict using GitHub’s AI assistant. If it passes, I repeat the 1-to-3 process for the next files.
- Inspect the code: From time to time, I ask OpenAI’s Codex in plan mode to run between one and four versions for inspection, depending on the number of tasks. For one version, I might write: “You are a technical reviewer responsible for auditing a website project for piano lessons. Review all the files in my repository, identify the most critical inconsistencies that could disrupt the workflow, and correct them.”
- Cross-check: Before merging, I copy and paste the issue along with Codex’s proposed fix into Claude Code (Sonnet 4.5 with extended thinking) for a context-aware second opinion. If Claude suggests a better solution, I feed it back into Codex for further refinement.
- Finalize the merge: Once new code is merged into GitHub’s main branch, a bug may still surface later in Codex’s code review section. Because Codex is connected in real time, it immediately flags the issue and proposes a correction for me to review and approve.
- Deploy: I use Vercel or Render, both connected to GitHub, to deploy the website and monitor its front-end performance.
Having tested a wide range of AI tools, I believe this is currently the most effective workflow for programming as of October 2025. I’m eager to keep exploring the next wave of features that will continue to shape how we build software in the months ahead.
Key Takeaways
- Vibe coding transcends small projects, with tech giants competing to master multi-file AI-augmented programming tools.
- Structured maintenance and smart input crafting result in significant time and cost savings in complex machine learning projects.
We tend to think vibe coding is only for people who can’t code, or that it’s limited to small projects. But it’s hard to talk about hindsight when the story only began earlier this year. Right now, OpenAI, Anthropic, Meta and Google are locked in fierce competition, investing billions of dollars to launch revolutionary AI tools that are all designed to handle multi-file projects.
As an engineer applying AI tools to complex projects, especially in machine learning, I’ve managed to cut both time and costs with three key solutions.
The rest of this article is locked.
Join Entrepreneur+ today for access.
