Generate Recipes from YouTube Videos with AI

This project follows the Engineering Design Process. Confirm with your teacher if this is acceptable for your project, and review the steps before you begin.

Generating YouTube Transcripts

1. On YouTube, find 10 recipe videos that you would like to generate recipes from.
   1. Make sure the videos include spoken narration, since we will use their transcripts to create the recipe.
2. Open the video description and scroll down to the bottom. You will see a button called “Show transcript.” If you click on that button, a new box with the transcript will appear.
   1. To get rid of the time stamps, click on the three dots on the top right corner of the transcript box and then select “Toggle timestamps.”
   2. Note: Not all YouTube videos will have transcripts.
3. Copy and paste the transcript into a text editor like Google Docs or Notepad. Give each separate transcript its own title or header in the document, or save each one in a different file. 
   1. Repeat step 4 for each of the 10 videos.

Experimenting with Recipe Generation and Prompt Engineering

1. Watch one of the videos you selected yourself and write down the recipe.
2. Think about the criteria you will use to evaluate the AI-generated recipe. For example:
   1. Does the recipe have any errors in ingredient quantities? If so, how many errors?
   2. Does the recipe have any steps out of order? If so, how many?
   3. Does the recipe have any hallucinations or seemingly fabricated information that cannot be verified?
   4. Is the recipe's format easy to follow? Is the language easy to read?
3. Create a data table that you can use to evaluate the recipes based on your specific criteria.
   
   Swipe left to see more

   Table 1. Example data table. You can add more rows to the table for additional videos and new prompts.

   Video	Prompt	Criteria #1	Criteria #2	Criteria #3	...	Notes

4. Generate a recipe from the video transcript using a basic prompt and evaluate the recipe.
   1. Choose an LLM like ChatGPT.
   2. Enter a prompt like “Make a recipe from this video transcript.” and then paste in the video transcript.
   3. Record your notes and observations in the data table you created.
5. Repeat step 4 for your remaining videos. Make sure you use the same LLM and prompt each time.
6. Now, this is where prompt engineering starts. Try modifying or adding something to your prompt. Here are several ideas to help you start, but you should only change one thing at a time.
   1. “Add ingredient alternatives to make this recipe vegan."
   2. “Include a difficulty rating.”
   3. “Add estimated preparation and cooking times.”
   4. “Make the recipe more detailed for someone who is inexperienced in cooking.”
   5. “Make the recipe more concise for someone who is experienced in cooking.”
   6. “Write the recipe so that the ingredients and quantities are built into the steps instead of having a separate ingredient list before the recipe.”
7. Repeat step 4 with your new prompt. If necessary, you should also rewrite the recipes yourself. For example, if you are going to prompt the AI to write a detailed recipe for an inexperienced chef, you should write the recipes in that style yourself so you can compare the results.
8. Continue to iterate and experiment with different prompts.
   1. How accurate is the LLM overall in determining the correct basic recipe based on the video transcript?
   2. What prompt gives you the best results for the format and style of a text recipe?