How I convinced our devs to use AI for coding (system prompt)

Embracing AI-Assisted Coding: How I Persuaded Our Development Team to Leverage AI Tools

In the fast-paced environment of a rapidly scaling startup, staying ahead of development workflows is crucial. Recently, I embarked on a journey to integrate artificial intelligence (AI) into our coding processes—a move that initially met with skepticism from our experienced backend engineering team. Here, I share the strategies I employed to foster trust and demonstrate the tangible benefits of AI-assisted development.

Understanding the Initial Resistance

Our team comprises dedicated backend engineers who take pride in their craftsmanship. When first introduced to AI coding tools, their immediate response was one of skepticism. They felt that the AI’s suggestions did not align with their coding style or standards, leading to doubts about its reliability. Consequently, the usage of AI remained minimal, hindering us from exploring its potential advantages.

Crafting a Trust-Building Strategy

To change their perspective, I recognized the importance of data-driven insights and creative problem-solving. My approach involved analyzing our historical development data—specifically, the last 500 pull requests (PRs). By examining the comments, observations, and structural patterns within these PRs, I aimed to identify consistent coding practices and areas where AI could add value.

Using Past Data to Inform AI Prompts

I took the insights gleaned from this analysis and employed multiple AI models to generate recommendations. These models identified key observations and crafted guidelines that would be most beneficial for onboarding new team members—essentially, creating a set of “best practices” distilled from our actual code history. These instructions underpin the development process and serve as a foundation for AI to produce more tailored and trustworthy code suggestions.

Iterative Testing with AI-Generated Code

Armed with these customized instructions, I used AI tools such as Claude Code or Cursor to generate initial code drafts for ongoing issues. These models were prompted with our established guidelines, leading to PR drafts that closely matched our standards. The results exceeded expectations: the initial AI-generated PRs were roughly 80% correct, allowing our engineers to focus on refinements rather than starting from scratch.

Results and Lessons Learned

The team’s immediate reaction was enthusiastic—remarking that the AI outputs were significantly closer to our expectations than before. This demonstrable improvement built their confidence in the tools, facilitating more frequent and effective AI-assisted coding.

Final Thoughts

The key takeaway from this experience is that integrating AI into a development workflow requires creativity, data analysis, and patience