Siddhant Khare, a software engineer at ONA and the lead maintainer of the OpenFGA project, recently published an essay that has resonated throughout the developer community. In the piece titled "AI Fatigue Is Real and Nobody Is Talking About It," he describes a paradox: AI tools are making him more productive than ever before, but they are also exhausting him more than ever before.
The Productivity Paradox
Last quarter, Khare shipped more code than during any other period of his career. At the same time, however, he felt more exhausted than ever before. AI really does speed up individual tasks—what used to take three hours now takes 45 minutes. The problem is that when every task takes less time, you do not do fewer tasks, but more of them. Before AI, he could spend an entire day on a single design problem. Now he can tackle six different problems in one day. Each one "only takes an hour with AI," but switching contexts between six problems is brutally demanding on the human brain. AI does not get tired when switching between problems. Humans do.
Before AI, Khare's work was clear: think about a problem, write the code, test it, ship it. With the advent of AI, his work has changed: write a prompt, wait, read the output, evaluate it, fix the parts that do not fit, prompt again, repeat. He has become an inspector on an assembly line that never stops. Creating gives you energy. Reviewing drains it. Khare first noticed this during a week of intensive AI use. By Wednesday, he could no longer make simple decisions. His brain was full. Not from writing code, but from judging code. Hundreds of small decisions, all day, every day. The cruel irony is that AI-generated code requires more careful review than human-written code. With AI, every line is suspect. The code looks confident, but it may be subtly wrong in ways that only become apparent in production, under load, at three in the morning.
Khare had a prompt that worked perfectly on Monday. On Tuesday, he used the same prompt for a similar endpoint. The output was structurally different. Why? For no reason he could determine. For someone whose career is built on "if something breaks, I can figure out why," this is deeply unsettling.
Write Just One More Prompt
Another problem is the pace of innovation. Claude Code ships sub-agents, OpenAI launches GPT-5.3-Codex, Google releases Gemini CLI, GitHub adds the MCP Registry. That is not one year. That is just a few months. Khare spent weekends evaluating new tools, watched every demo, and feared falling behind. Each migration cost him a weekend and delivered perhaps a 5% improvement that he could not even properly measure.
You are trying to get AI to generate something specific. The first output is 70% correct. You refine the prompt. The second output is 75% correct, but it breaks something the first one got right. You have already spent 45 minutes on it, when you could have written it from scratch in 20 minutes. Khare now has a strict rule: three attempts. If AI does not produce something 70% usable within three prompts, he writes it himself. No exceptions.
During a design meeting, someone asked him to work through a concurrency problem on a whiteboard. No laptop. No AI. And he struggled. Not because he did not understand the concepts, but because he had not exercised that muscle for months. He had outsourced his initial thinking to AI for so long that his ability to think from scratch had deteriorated. It is like GPS and navigation. After years of using GPS, you cannot navigate without it. The same thing is happening with AI and engineering thinking.
What Helped Him?
Khare changed his approach:
Time-limiting AI sessions—30 minutes for a task with AI. When the timer runs out, he ships what he has or starts writing it himself.
Separating time—mornings are for thinking, afternoons for execution with AI assistance.
Accepting 70% from AI—he stopped trying to get a perfect output. He fixes the rest himself.
Taking a strategic approach to the hype cycle—he stopped adopting every new tool a week after launch.
The Real Skill
Khare burned out in late 2025. Not dramatically—he simply stopped caring. He went through the motions, produced more than ever before, and felt less than ever before. The tech industry had a burnout problem long before AI. AI is making it worse, not better. Before AI, there was a ceiling on how much you could produce in a day. AI removed that limit. Now the only constraint is your cognitive stamina.
The real skill of the AI era is not prompt engineering or knowing which model to use. It is knowing when to stop. Knowing when AI's output is good enough. Knowing when to write it yourself. Knowing when to take a break or stop. If you are tired, it is not because you are doing it wrong. It is because this is genuinely hard.
