The text comprehension and generation capabilities of large language models (LLMs) have become extremely powerful, fulfilling the needs of many application scenarios. And online chatting is very popular, which provides new application scenarios for text steganography. In this paper, we propose a steganography scheme that generates conversational text with specified topics and grammars based on the LLMs. The algorithm maps a secret as prompts input into the LLMs to generate steganographic text. To improve imperceptibility, after statistical analysis of tense usage in different topics, its frequencies are used to code the tenses in variable lengths to ensure that the generated steganographic text aligns the tense of the generated steganographic text with that of normal text. The comprehensive analyses are performed to evaluate the generated steganographic text. The results show that in term of Perplexity, a steganographic text has good imperceptibility, and the metrics such as Kullback-Leibler divergence (KL) and Jensen-Shannon divergence (JS) indicate it is statistically difficult to distinguish the generated steganographic text from normal text, which indicates the steganographic text has strong anti-steganalysis ability.