Posts

AI can write tests in seconds, but coverage numbers are lying to you. Here’s how tautological, implementation-coupled tests let a real bug survive for three weeks.

AI tools can genuinely help with legacy code — but only on the right problems. Understanding and documentation: yes. Refactoring core logic you don’t fully understand: hard no.

An AI-generated migration ran perfectly in dev. In production, it would have locked a large production table during business hours. What I learned the hard way about AI and database changes.

Senior engineers have seen enough hype cycles to be skeptical of AI tools — and their concerns about maintainability and debugging complexity are legitimate. But refusing to engage is its own kind of risk.

Codex just hit 1M weekly active users, GPT-5.4 under the hood, Figma MCP integration. Claude Code has been my daily driver for months. I gave both the exact same tasks for a week and tracked speed, code quality, context handling, pricing, and GitHub integration. Neither won cleanly.

The NYT’s ‘Coding After Coders’ piece made a lot of engineers defensive. As a Tech Lead managing a small engineering team, I read it twice — what they nailed, what they completely missed, and why the real story is more complicated than either side admits.

AI coding tools quietly moved from flat subscriptions to credit-based pricing. I tracked my small team’s actual spending for a quarter and found we were paying well over $300/month per developer instead of the advertised $20. Here’s the breakdown, the traps, and how we cut costs by 40% without losing productivity.

In 3 months my team evaluated 6 AI coding tools, switched primary tools twice, and lost an estimated over 100 engineer-hours to setup, configuration, and relearning. I finally enforced a 90-day moratorium on tool changes. Here’s what I learned about the real cost of chasing the next shiny AI tool.

Cursor launched Automations on March 5, letting you set up always-on agents triggered by code changes, Slack messages, or PagerDuty alerts. After a week of testing, here’s what it actually looks like to manage a team where agents review every PR before a human even opens it.

GPT-5.4 launched March 5 with native computer use and integrated coding from GPT-5.3-Codex. Here’s what three days of real coding work revealed — the genuinely impressive parts, the benchmarks that don’t tell the full story, and why I’m still keeping my Claude subscription.