A Towards Data Science piece shows how the author turned ad‑hoc LLM prompting into a repeatable research pipeline using a Claude Code skill. While ChatGPT or Claude queries produce fast answers, they drift across sessions and resist automation. By wrapping persona generation, panel design and answer collection into a single command, the workflow gains consistency without sacrificing the natural‑language flexibility that makes LLMs useful.

The author demonstrates the approach with virtual customer research, generating ten Gen Z skincare shoppers in the US via `/persona generate 10 Gen Z skincare shoppers in the US`. Each persona is stored as a JSON object, preserving attributes like age, attitude and budget tier. Pre‑defining panel diversity and validating the distribution prevents the drift that plagues ad‑hoc chats, enabling reliable follow‑up queries.

Unlike a pure Python library such as Microsoft Research’s TinyTroupe, the Claude Code skill runs inside the existing Claude subscription, avoiding extra API keys and billing friction. Parameters arrive as natural language, and a SKILL.md file guards the workflow structure, so users can reuse the same panel for new concept tests. This hybrid of code and prompt delivers a scalable, reproducible method for qualitative AI‑driven research.