Quick tip here: Where possible try to trick the model into thinking whatever task it is writing will / could be read and performed by one of your coworkers rather than another agent. It will quickly fall into line and stop writing prompts meant for dumb historic low intelligence LLMs.