One thing that's helped me a little bit, is to open up the spec as context, and then asking the LLM to generate tests based on that spec.