Using Codex Is a Lot Like Baking (elijahpotter.dev)

A veteran developer reports that Codex-style models are already capable of economically useful web development work—understanding codebases, making targeted changes, and iterating until a prompt is satisfied—but the main shortcoming isn’t raw ability: it’s the common workflow. Instead of proactively asking clarifying questions and building an internal plan, many agents rely on a back-and-forth “vibe-coding” loop that forces a human to micromanage iterations. In an experiment, the author wrote a concise project-style goal document (what/why/how plus anticipated edge cases), handed it to Codex, and returned an hour later to find the landing page changes complete—far more efficient than interactive prompting. Technically, this highlights that LLMs function well as autonomous batch workers when given structured, unambiguous specifications. The practical implication for AI/ML teams is to invest in richer prompt engineering: treat agents like contractors by encoding requirements, constraints, and expected outputs up front, rather than relying on conversational clarification. It also exposes a model-level gap—lack of proactive question-asking—that could be addressed by tooling that extracts intent, requests missing info, or supports goal documents as first-class inputs. For engineers, the takeaway is simple and actionable: write better project docs for your LLMs and let them “bake.”