To mitigate the "cognitive debt" incurred by over-reliance on Large Language Models (LLMs), the author proposes a workflow centered on manual code transcription rather than direct copy-pasting. Instead of allowing AI agents to submit pull requests or automate entire features—a process the author describes as reviewing "subtly incorrect" or "badly-commented" code—they request LLM-generated code within a chat interface and manually retype every line into the editor. This method prioritizes comprehension over raw output speed; while traditional LLM usage might offer a 10x speed increase, this manual approach yields only approximately a 2x speedup. However, the author argues this trade-off is essential for building a robust mental model and a "spatial map" of the codebase. By slowing down the implementation process, the developer can more effectively detect hallucinations, refactor logic in real-time, and ensure a deep understanding of APIs and algorithms. This deliberate approach aims to prevent a future where developers lack fundamental understanding of the digital infrastructure they maintain, treating code comprehension as a safeguard against professional malpractice in an era of increasing automation.