The Prompt-Wait-Evaluate Loop: How AI Kills Flow Without You Noticing | Sandor Dargo's Blog<br>Sandor Dargo's Blog<br>On C++, software development and books

HOME TAGS ARCHIVES BOOKS SPEAKING DAILY C++ WORKSHOPS HI... SUBSCRIBE

Blog 2026 07 15 The Prompt-Wait-Evaluate Loop: How AI Kills Flow Without You Noticing Post<br>Cancel

The Prompt-Wait-Evaluate Loop: How AI Kills Flow Without You Noticing<br>Sandor Dargo Jul 15 2026-07-15T00:00:00+02:00<br>11 min

A few months ago, I wrote about finding joy in programming in the age of AI. In the personal discussions I’ve had with fellow developers — both before and after that article — one thing keeps striking me. Most people don’t argue. They just say: yes, that’s exactly how it feels.<br>But a question kept coming up in those conversations: why does it feel this way? Not in the philosophical sense — I covered that. In the mechanical sense. What is actually happening to our attention when we work with AI coding assistants?<br>I want to explore that topic more deeply, because I think understanding the mechanism matters. It’s the difference between vaguely sensing that something is off and being able to do something about it.<br>What flow actually is<br>In 1990, the psychologist Mihaly Csikszentmihalyi published Flow: The Psychology of Optimal Experience, the book that gave a name to what probably all of us had already experienced. Flow is the state of full absorption where action and awareness merge. You lose track of time. Self-consciousness fades. The work just happens.<br>Csikszentmihalyi identified three preconditions:<br>Clear goals. You know what you’re trying to accomplish at each moment, not just at a high level, but at the resolution of the next few keystrokes.Immediate feedback. You can tell right away whether what you just did worked. The compiler, the test, the output — something tells you where you stand.Challenge that matches your skill. The task is hard enough to require your full engagement but not so hard that you freeze. You’re stretched, not stuck.Programming, at its best, is one of the purest flow states there is. That’s half of why we loved it. Think about a good debugging session. You have a clear goal — find the bug. You get immediate feedback — the test passes or it doesn’t, the log says this or that. And the challenge scales with you. As you eliminate hypotheses, the search narrows. You’re fully in it.<br>The same goes for writing a new component from scratch. You hold the architecture in your head, you lay down the skeleton, you iterate. Each compilation tells you whether you’re on track. The difficulty is real but manageable. Hours vanish.<br>I think most experienced developers can point to moments like this. These are the sessions you remember fondly, even years later. You were not just productive. You were alive in the work.<br>The old thieves<br>Flow has always had enemies. We already knew them well before AI entered the picture.<br>Meetings that break your morning in half. Slack messages that make you look away for just a second. Email notifications. Code review requests that pull you out of your current task and into someone else’s context. Somebody popping up at your table in the open office just to have you for a minute. The daily standup scheduled at 11 AM, right when you were about to get into it.<br>Gloria Mark, a professor at UC Irvine who has spent over two decades studying how knowledge workers actually spend their time, found that after a single interruption it takes an average of 23 minutes and 15 seconds to return to the same level of focus. Not to start the same task again — to reach the same depth of engagement. Her follow-up research, published in her 2023 book Attention Span, showed that the average time people spend on a single screen before switching has dropped from 2.5 minutes in 2004 to just 47 seconds.<br>These numbers are not new. We have known about them for a long time. But they matter now more than ever.<br>Because AI added a new thief. And it’s sneakier than all the old ones.<br>The new interruption you invited in<br>Here is what a typical AI-assisted coding session looks like.<br>You have a task. You think about it for a few moments. You write a prompt. You submit it. You wait. The model generates something. You read it. You evaluate it. Maybe it’s close, maybe it’s off. You adjust the prompt. You submit again. You wait again. You read again. The cycle repeats.<br>Prompt. Wait. Evaluate. Re-prompt.<br>This is not flow. This is the opposite of flow. Let me walk through why.<br>The goals are no longer clear at each moment. When you type code yourself, you know what the next line should do. You hold a mental model and you execute against it. When you write a prompt, you’re describing what you want at a high level — but you have no moment-to-moment clarity about what the machine will produce. You are waiting for a surprise.<br>The feedback is no longer immediate. Instead of the tight loop of type-compile-see, you get a chunk of output after a delay. With a chat-based assistant, that...