In many robotic applications, robots operate in shared workspace where users may modify or supervise tasks. In Small and Medium Enterprises, the adoption of automation is still limited by the necessary robotic skills required for programming manipulators. An idea is to give focus on Large Language Models (LLMs), even if the lack of trust from the human side represents a significant obstacle. This thesis presents a “user on the loop” approach in which an LLM interprets a high-level robotic assistant. From a natural language request, the model generates a pseudocode that is validated through simulation, refined through feedback iterations, and executed on the real hardware by the user.