Abstract <p>Growing possibilities of modern large language models discover the way to create new system for reasonable control of robot behavior using a limited communication with an operator. The system is proposed to convert human instructions into code snippets taking into account information from the robot’s world. This approach translates the complex interaction logic of the responses of the large language model into a code execution environment, resulting in improved reliability and predictability of system behavior. This system was evaluated using the set of handwritten instructions. This set consists of three parts: ‘‘direct’’ instructions relate to objects actually present in the environment, ‘‘abstract’’ instructions describe general tasks without specific details, and ‘‘out-of-space’’ instructions refer to actions beyond the robot’s capabilities or mention objects not shown in its environment. Large language models, such as Deepseek, Qwen, and GPT, were tested with the different types of training and the number of parameters. The results demonstrate that generating control instructions in the form of executable code increases the flexibility of robot control and reduces the risk of hallucinations, particularly during arithmetic operations.</p>