Human-in-the-Loop Software: From Prompts to Pull Requests

Disclaimer: The views and opinions expressed in this article are my own and do not necessarily reflect those of my employer or any affiliated organizations.

If you've been using generative AI over the past year, you may have found yourself "vibing" your way through an Excel formula or creating an interactive flashcard game for your kids.  Vibe coding, a term for building software by describing what you want and letting AI generate the code is gaining traction this year, despite skepticism. Critics argue it churns out sloppy and unmanageable code, particularly when used by novice coders, but software development is undeniably being transformed.

Prompt engineering defined how we talked to AI in 2024; vibe coding defines how we're building with it in 2025

AI-Assisted Software Development

AI tools accelerate prototyping, automate mundane tasks, and even suggest novel solutions to common problems. But if left unchecked, these tools can produce inefficient or convoluted code, a nightmare for debugging and scalability.

Vibe coding can deliver results quickly, but without the right fundamentals, it can waste precious engineering time. The best of AI-assisted software development is an iterative partnership between human oversight and machine-generated code. Just like in traditional software development, clearly defining requirements and specifications remains essential.

According to a 2023 McKinsey report, developers using AI tools like GitHub Copilot can code up to 56% faster. But speed means nothing without structure. AI can deceptively produce functional yet messy code. The trick is treating AI-generated code exactly like human-written code. Our processes still include code reviews, pull requests, automated testing, and thoughtful documentation. Ensuring every commit is reviewed and refined through human oversight transforms AI from a reckless contributor into a powerful collaborator.

Real-World Impact

I've heard a lot about how large language models (LLMs) can turn a beginner developer into a junior, and a junior into a senior. Of course, engineering judgment and domain expertise are irreplaceable and AI will augment, not replace, the developer’s role. These tools remove barriers to entry, surface best practices in real time, and give developers at any stage a kind of instant mentorship.

In 2024, Google began incorporating its AI assistant, Codey, into internal development environments to automate repetitive tasks and assist in large-scale refactoring. According to Google’s own research, shared during the Google I/O keynote, developer satisfaction and code quality metrics improved significantly when AI tools were used alongside pair programming and Agile processes.

In 2025, we’re seeing an emerging class of developers experimenting with automation pipelines. Take Carbon CLI. In a demonstration on YouTube, Ovadiah Myrgorod shows how to use ChatGPT to autonomously generate and commit new code based on GitHub issues. Codebases are already evolving semi-autonomously, reshaping how developers work.

Git and PRs: Human-in-the-loop

Tools and processes like Git and PRs keep AI-accelerated development tethered to engineering reality. Git provides version control, accountability, and traceability of AI-generated changes, while PRs allow for essential human scrutiny and alignment with team quality standards.

With this new agentic workflow, traditional techniques like test-driven development and behavior-driven development are elevated. Comp Science principles like extensibility, DRY and object-oriented design continue to underpin well-architected software. While vibe coding may sound casual, it’s not a shortcut, it’s a new interface to an old discipline.

With generative AI, extensibility today is just as much a workflow and collaboration issue as it is a technical one. It’s no longer just about writing flexible code, but also about designing systems that anticipate interactions between humans and machines.

As LLMs increasingly contribute to active codebases, extensibility means designing for iteration. That means writing clear, testable code, with modular responsibilities and explicit behaviors, so that future enhancements remain manageable.

This shift requires developers to think beyond current frameworks. Versioning, interface clarity, and semantic naming conventions become part of the architecture itself. In many ways, AI is encouraging better habits because any ambiguity gets amplified at machine scale. Teams that embrace this shift and design for adaptability will extract compounding value from AI-assisted development.

Vibe Check

In 2025, automation isn’t just a trend, it’s a shift that has playing out for some time. AI isn’t confined to code editors or dev environments anymore. It’s redefining productivity across the enterprise. From writing smarter Excel formulas, to marketers generating landing pages with no-code tools, to cybersecurity teams automating detection and response scripts, coding is no longer a siloed discipline. It’s becoming a core competency across the board.

While professional software engineering underpins the broader wave of intelligent automation, coding is welcoming more people, across more roles, into the world of software creation.

Chui, Michael, et al. The Economic Potential of Generative AI: The Next Productivity Frontier. McKinsey & Company, 14 June 2023, https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier.

Google. (2024). Google I/O Keynote. Retrieved from https://io.google (specific keynote URL needed).

Myrgorod, O. (2025). Project Carbon: AI-powered development automation with ChatGPT + GitHub Actions [Video]. YouTube. https://www.youtube.com/watch?v=jfB6wiBvpFQ.