Vibe-coding is destroying my work-life balance

I’ve hit a wall where my productivity is actually becoming a liability to my mental health. It’s a weird paradox: I’m technically getting more done, but I’m losing the thread of my own codebase. I’ve fallen into a cycle of "vibe-coding" with Claude Code, and it’s turning my professional engineering workflow into a high-speed, low-comprehension loop that keeps me glued to my desk long after I should have logged off.

The shift is subtle. In a healthy LLM-assisted workflow, I’m the architect. I design the logic, write the core structures, and use Claude Code as a high-powered pair programmer to handle boilerplate or solve specific, isolated logic hurdles. I stay in control.

But as the day progresses and fatigue sets in, my discipline slips. I stop being the architect and start being a spectator. I enter this state where I’m just prompting, hitting "accept," and watching the terminal fly. I’m not even reviewing the diffs properly anymore; I’m just riding the wave of the LLM's momentum. This "vibe-coding" feels deceptively easy because the cognitive load of actual problem-solving has been outsourced, so my brain doesn't signal that it's "working hard." Consequently, I don't feel the natural exhaustion that usually tells me to stop. I just keep pushing "one more feature" until it's 10 PM and I'm staring at a massive, unreviewable chunk of code that I don't fully understand.

I’ve tried a few tactical shifts to regain control of my AI workflow, but nothing is a perfect fix:

• Manual Diff Reviewing: I’ve tried forcing myself to manually type out the edits Claude suggests instead of using the "auto-apply" or "accept" features. It forces a character-by-character review that keeps me grounded in the logic. It works, but it’s incredibly draining, and I usually abandon it the second I feel that late-night urge to just "get it done."
• Granular Task Decomposition: I’m trying to break my tickets into much smaller, micro-chunks. The goal is to force a natural break in the flow—forcing a PR, waiting for a colleague's feedback, or jumping into a design discussion. It breaks the hypnotic loop of the terminal.
• Constraint-Based Prompting: I’ve actually tried telling the agent, "Hey, it's 6 PM, tell me to stop working," but in long, context-heavy sessions, the LLM loses that instruction and goes right back into "helpful assistant" mode.
• Strict Hard Deadlines: Setting a non-negotiable shutdown time. This is the hardest part because the "vibe" is so addictive.

The real issue might be deeper than just tool usage. When you work remotely, your primary collaborator shifts from human teammates to the LLM. The friction of human communication (Slack, meetings, PR reviews) used to act as a natural brake on my productivity. Without that friction, Claude Code provides a frictionless path to burnout. I’m essentially fighting against the very efficiency I sought to gain.