A new solution to the task of prompt recovery in text rewriting is presented in this paper, in which it is sought to fill in the missing rewrite prompt that transforms an original text into a rewritten one. The research employs a dataset of original and rewritten pairs of texts, which develops models that are designed to determine the rewrite prompt. The proposed pipeline consists of a reproducible system, data auditing, feature engineering on the differences between the original and rewritten text, and testing with such metrics as BERTScore F1, Word Overlap, BLEU, and ROUGE. We find that the Retrieval Baseline has the highest average BERTScore F1 of about 1.0000, which means that it is strongly semantically similar to the gold prompts. The Heuristic Baseline registered a BERTScore F1 of about 0.8501 and the lowest score of 0.8111 was recorded in the LLM Approach. Also, the BERTScore of the Heuristic and LLM methods was very high relative to the Word Overlap scores, and this indicates that BERTScore is capable of capturing more semantic similarity. Nevertheless, the prompts the LLM predicted had a very low Word Overlap of 0.0000, and the predicted text was the one that was actually rewritten, meaning that although there were certain semantic cues, the approach of the LLM was not effective in leading the text generation in this activity. The paper ends with reflections on the advantages and limitations of various methods to ensure prompt recovery and proposes directions to focus on in the future.