Long coding sessions with Claude and Codex can leave developers mentally exhausted, according to a detailed account from a software engineer. The author describes hitting a wall after 4-5 hour sessions where fatigue leads to sloppy prompts, slower feedback loops, and mounting frustration. Rather than blaming the models themselves, the engineer identifies patterns in their own workflow that create a downward spiral.

Poor prompting quality emerges as a key culprit. When mentally drained, developers rush prompts, interrupt the AI mid-response, and skip crucial context. This creates a feedback loop where the AI's responses worsen, prompting more frustration. The engineer also highlights technical bottlenecks - parsing large files takes minutes, causing context windows to fill up before experiments complete. This forces choices between dumbing down the AI or restarting sessions.

Solutions focus on recognizing personal limits and optimizing workflows. The engineer now stops when prompts feel rushed or uncertain, treating that as a signal to rest. For slow tasks, they've adopted a test-driven approach with the AI, setting clear success criteria and timing constraints to speed up iteration. The core insight: AI coding fatigue often stems from skill gaps rather than model limitations. By treating slow feedback as the problem to solve and maintaining prompt discipline, developers can avoid the exhaustion trap and work more effectively with LLMs.