Twelve years of professional web development. Daily access to an AI coding assistant. And the work is somehow harder to finish than it was before.
This isn't a complaint from someone who never got the hang of AI tools. It's the experience being reported by seasoned developers who have given these tools a genuine run and come back with mixed verdicts.
The specific mechanism: the waiting loop. You send a prompt to Claude Code, and while the model processes your request - typically 10 to 30 seconds for anything complex - you do what humans naturally do with a pause. You open another tab. You start a Slack message. You check something quick. Then the response lands and you're half-committed to two things instead of fully committed to one.
Cognitive psychologists call this "attention residue" - the mental overhead that lingers when you switch tasks before the first one is done. AI assistants haven't solved this problem. For some workflows, they've amplified it by introducing a new waiting cadence that breaks up what used to be uninterrupted stretches of work.
How Waiting Loops Fragment Deep Work
The developers who report the clearest productivity gains from AI describe narrow, specific use cases. Stuck on a regex you can't crack. Need to scaffold a route type you've never written. Want a second pass on a function's error handling. Discrete problems with a clear start and end.
The returns shrink fast when AI gets woven into the continuous flow of work - used as a running commentary on every decision rather than a tool you pick up for a specific job and put back down. At that point, the workflow becomes a prompt-response loop, and what gets lost is the sustained, single-track focus that complex technical problems actually require.
There's a subtler cost too: the gap between executing AI-generated code and understanding it. When a model handles the reasoning behind a decision, the developer implementing that decision skips the problem-solving process that builds expertise over time. Twelve years of experience is partly twelve years of getting stuck and wrestling through it. AI can short-circuit that in ways that feel like progress in the moment.
Narrower Is Usually Better
The strongest case for AI in developer workflows is still as a specialist tool, not a general copilot. Ask it specific questions when genuinely blocked. Review what it produces critically. Keep the long stretches of focused work outside its loop.
This isn't an argument against AI tools - they're clearly useful for the right jobs. The mistake is treating more AI as automatically more productivity. The honest read from experienced developers is that these tools work best when they stay in their lane: specific, bounded, and on request.
The productivity story around AI has always had a small-print clause: the tools work well for discrete tasks, but they haven't yet cracked how to support the sustained, deep work that complex projects depend on. Figuring out where to draw that line is now one of the more useful skills in software development.