My article on why AI is great (or terrible) or how to use it
Matthew Rocklin’s take is refreshingly practical: senior engineers are uniquely positioned to get real leverage from AI coding tools, because they can steer work at a higher level and detect when outputs are drifting into “slop.” He’s explicit about tradeoffs — LLMs generate junk (often lots of it), and the hardest part of shipping software is reviewing for correctness, not emitting code. The question, then, is how to keep the benefits without turning your day into approving one prompt after another.
Two of the core suggestions are (1) add structure around the agent, and (2) find ways to build confidence without line-by-line reading. On structure, he highlights practical tooling like Claude Code hooks to enforce conventions (e.g., how tests are run), smooth over permission prompts, and reduce the dehumanizing “yes, allow” loop. On confidence, he argues we need better habits to validate large AI-assisted changes: tests, smaller feedback loops, targeted review, and even using AI to assist with review, not just generation. It’s a useful framework for anyone trying to integrate coding agents into a serious engineering workflow without letting throughput outrun understanding.