# Getting Every ReviewBrah Rating Using LLMs

# "Hello Everyone!"

A phrase that echoes throughout my house on weekly basis from my favorite YouTuber, ReviewBrah.

# How I did it

The process includes retrieving transcripts, aligning them with upload dates, crafting a custom LLM prompt, and performing a final cleanup pass.

# Step 1: Installing and Using youtube-transcript

Install the youtube-transcript tool:

cargo install youtube-transcript

I then tested the installation by downloading a single transcript:

youtube-transcript "https://www.youtube.com/watch?v=V4PO9nsW3aU" > reviewbrah2.txt

At this point, reviewbrah2.txt contains raw transcript text of the single video.

# Step 2: Listing All Videos in a Channel

I used yt-dlp to retrieve a list of videos from the channel

yt-dlp -s --flat-playlist --print-to-file url list.txt "https://www.youtube.com/@TheReportOfTheWeek/videos"

Your list.txt might look like:

https://www.youtube.com/watch?v=w_fCkwqW0DM
https://www.youtube.com/watch?v=Rys_CRK76sg
https://www.youtube.com/watch?v=V4PO9nsW3aU
...

Each line includes only the corresponding video URL.

# Step 3: Downloading Transcripts for Each Video

Use a loop to process every URL in list.txt and download the corresponding transcripts:

while read -r date url; do
video_id=$(echo "$url" | awk -F'v=' '{print $2}')
    youtube-transcript "$url" > "${video_id}\_output.txt"
done < list.txt

Now you have \*\_output.txt transcript for each video.

# Step 4: Prepending the Upload Date to Each Transcript

Next, prepend the corresponding upload date to the top of each transcript file:

I noticed that the transcripts didn't have a date in the files, so I had to go back to retrieve them, then annotate each transcript.

First, get the dates for each transcript:

yt-dlp --skip-download \\n  --print-to-file "%(upload_date)s %(webpage_url)s" list.txt \\n  "https://www.youtube.com/@TheReportOfTheWeek/videos"

Then prepend each transcript with the correct date

while read -r date url; do
video_id=$(echo "$url" | awk -F'v=' '{print $2}')
    {
        echo "$date"
cat "${video_id}_output.txt"
    } > temp && mv temp "${video_id}\_output.txt"
done < list.txt

Now each transcript starts with its upload date

# Step 5: Crafting a Custom LLM Prompt

The LLM prompt must be precise so the model outputs only a markdown table. You may need to experiment to remove headers or other extraneous text. For example:

llm --system "Output only a markdown table with rows of data for columns: Reviewed Item, Restaurant, Cost, Rating, Review Date. Do not include a header row or any text other than the table rows themselves." --save reviewbrah

I had to adjust this prompt a bit to get the desired output format, and even then there would be some inconsistencies from time to time.

# Step 6: Feeding Transcripts into the LLM

Combine the transcript files with list.txt and feed them into the LLM using the refined prompt:

while read -r date url; do
video_id=$(echo "$url" | awk -F'v=' '{print $2}')
    files-to-prompt "${video_id}\_output.txt" | llm -t reviewbrah
done < list.txt | tee -a combined_output.txt

At this stage, you have a combined table spanning multiple videos, but it may still need cleanup due to inconsistencies in llm output, but just use llm again to clean up its mess.

# Step 7: Cleaning and Normalizing the Final Table

Perform a final cleanup step to ensure dates follow YYYY-MM-DD format and that no extraneous text remains:

cat combined_output.txt | llm --system "Clean this output up and make a singular markdown file from it, ensure all dates follow YYYY-MM-DD format. Only output the markdown table." > table.md

table.md now contains a unified markdown table. All dates are normalized, and the table is ready to publish.

# Conclusion

By combining: • youtube-transcript to extract text • yt-dlp to get metadata and dates at a channel level • Shell scripting to align transcripts with their upload dates • A carefully crafted LLM prompt to ensure the correct table format • A final cleanup pass with the LLM

…you produce a neat, organized markdown table.