It is difficult to find a recipe that uses the ingredients in a person’s refrigerator within a short time. To solve this problem, we propose a recipe–generation model in the encoder– decoder framework. Models developed in the traditional encoder– decoder framework do not adequately reflect the ingredients in cooking recipes, but the proposed method introduces reinforcement learning and coverage loss. The model was experimentally evaluated on a dataset of approximately 15 K cooking recipes extracted from Food.com. The evaluation index was ingredient matching (IM), a new evaluation metric, showing the extent to which the recipe uses the input ingredients. Relative to the existing model, the proposed model improved the IM by approximately 21%.