The writer reports daily interaction with LLMs, using Claude Code at work and Codex at home. They follow a single‑task workflow, letting the assistant generate code but reviewing and revising it themselves. This practice exposes them to novel approaches and fills knowledge gaps. Their role has shifted from writing code to designing it, feeding the design into an LLM, inspecting the output, and finalizing the implementation.

Their current focus is building a framework for large‑scale, unsupervised code generation within their codebase. They sift through outputs from Qwen, an unsupervised agent, while also creating tooling around Claude. For quick queries they rely on Chat GPT or Gemini’s overview, falling back to browsing only when the model fails. This hybrid workflow has lasted a year and the writer reports increased productivity.

However, repetitive writing patterns—false assumptions, hallucinations, staccato fragments, emoji overload—have begun to erode enjoyment. Similar mistakes from different models create a fatigue loop. Personalization features help but do not eliminate style repetition. The writer’s frustration highlights a need for tools that adapt tone, reduce hallucinations, and provide better control over generated style.