AI Journal 1: A Developer's Slow Surrender to AI

Summary

This first-person journal traces an experienced software developer's changing relationship with AI coding tools from the arrival of ChatGPT through the rise of agentic coding. While teaching software development, the author used ChatGPT for demonstrations and simple exercises, but found that it failed on a difficult XML and Java client task. After returning to professional development, the author used ChatGPT for documentation, configuration, code discussion, and architectural brainstorming, while retaining responsibility for the actual work. GitHub Copilot then accelerated code completion and especially unit-test writing; after the author presented its benefits, the company purchased licenses for roughly 20 employees. As AI tools became normal at work, the author concluded that access to them did not make inexperienced developers equal to skilled ones, because judgment and experience still shaped the results. The author initially rejected “vibe coding” and assumed AI could not produce code as well as an experienced human. That view changed after experimenting with OpenAI Codex and later Claude Code, which could inspect files, suggest changes, and perform edits with permission. The author now sees agentic tools as genuinely useful and difficult to abandon, while remaining uneasy about what widespread delegation of coding work may mean for software development.